Cognition on Cognition

< BACK

Cognition on Cognition Edited by Jacques Mehler and Susana Franck

Preface: Building COGNITION I

Neuropsychology

1 Insensitivity to future consequences following damage to prefrontal cortex 2 Autism: beyond "theory of mind" 3 Developmental dyslexia and animal studies: at the interface between cognition and neurology 4 Foraging for brain stimulation: toward a neurobiology of computation November 1 9 9 5 ISBN 0 - 2 6 2 - 6 3 1 6 7 - 9 504 pp. $ 5 5 . 0 0 / £ 3 5 . 9 5 (PAPER)

5 Beyond intuition and instinct blindness: toward an evolutionary rigorous cognitive science II Thinking 6 Why should we abandon the mental logic hypothesis? 7 Concepts: a potboiler

ADD TO CART

Series Bradford Books Cognition Special Issue Related Links Contributor List Request Exam/Desk Copy

8 Young children's naive theory of biology 9 Mental models and probabilistic thinking 10 Pretending and believing: issues in the theory of ToMM 11 Extracting the coherent core of human probability judgment: a research program for cognitive psychology 12 Levels of causal understanding in chimpanzees and children 13 Uncertainty and the difficulty of thinking through disjunctions III Language and Perception. 14 The perception of rhythm in spoken and written language 15 Categorization in early infancy and the continuity of development 16 Do speakers have access to a mental syllabary? 17 On the internal structure of phonetic categories: a progress report 18 Perception and awareness in phonological processing: the case of the phoneme 19 Ever since language and learning: afterthoughts on the Piaget-Chomsky debate 20 Some primitive mechanisms of spatial attention 21 Language and connectionism: the developing interface 22 Initial knowledge: six suggestions

Preface: Building COGNITION The human mind needs to acknowledge and celebrate anniversaries; however, some anniversaries are more salient than others. This book emanates from Volume 50 of the journal, COGNITION. Why that volume of COGNITION was important to us perhaps becomes clear when we understand how the mind encodes numbers. Indeed, Dehaene et al. (1992) reported that the number 50 is psychologically more salient than, say, either 47 or 53. So, predictably, Volume 50 was a befitting occasion to celebrate an anniversary; it was a time to take stock of what was happening during the early years and a time to remember how we were long ago and how we have evolved as a journal. In our first editorial, we wanted to remember those who have provided us with so much help and the cultural climate that made the journal possible. In this introduction to COGNITION on Cognition we leave as much of the original introduction as possible so that the flavor initially conveyed remains. COGNITION was envisioned by T. G. Bever and Jacques Mehler because we thought that the new and diffuse area of cognition had to be facilitated by overcoming the inflexibility of form and content that were characteristic of most earlier journals in psychology and linguistics. Moreover, cognition was a multidisciplinary domain while psychology and linguistics were too narrow and too attached to one school of thought or another. So too were most journals. In the sixties, one could see the birth of the cognitive revolution in Cambridge, Massachusetts, where many of those who were to become the main actors were working on a project which was to become modern Cognitive Science. Was it possible to study intelligent behavior, in man and in machine, in the way that one studies chemistry, biology or even astronomy? We were sure the question should be answered affirmatively. Since then, the study of mind has become a part of the natural sciences. Positivism and behaviorism, among others, had confined publishing to patterns that were ill-suited to our needs. Psychologists, linguists, neuropsychologists, and others would often voice their dismay. Authors knew that to enhance their chances of publication they had to avoid motivating their studies theoretically. "Make your introduction as short and vacuous as possible" seemed to be the unspoken guideline of most journals. Editors were often even more hostile towards discussions that had "too much theory," as they used to say in those days. That was not all. Psychology journals did not welcome articles from linguistics while neuropsychologists had to hassle with neurologists to see their findings published. For a psychologist to

VIII

Preface

publish in a linguistics journal was equally out of bounds. Readership was also broken down along lines of narrow professional affinity. Yet scientists from all these disciplines would meet and discuss a range of exciting new issues in the seminars held at the Harvard Center of Cognitive Studies, and at similar centers that were being created at MIT, Penn, amongst others. Those were the days when computer scientists and psychologists, neurologists and linguists were searching jointly for explanations to the phenomena that their predecessors had explored from much narrower perspectives. If perception continued to be important, learning was beginning to loose its grip on psychology. Neuropsychology and psycholinguistics were becoming very fashionable and so was the simulation of complex behavior. Studying infants and young children had once more become a central aspect of our concerns. Likewise, students of animal behavior were discovering all kinds of surprising aptitudes to which psychologists had been blinded by behaviorism. It was, however, in the fields of linguistics and computer science that the novel theoretical perspectives were being laid out with greatest clarity. What was wanted was a journal that could help students to become equally familiar with biological findings, advances in computer science, and psychological and linguistic discoveries, while allowing them to become philosophically sophisticated. So, some of us set out to create a journal which would enclose such a variegated domain. We also wanted a publication for which it would be fun to write and which would be great to read. These ideas were entertained at the end of the sixties, a difficult time. France was still searching for itself in the midst of unrest, still searching for its soul after hesitating for so long about the need to face up to its contradictions, those that had plunged it into defeat, occupation and then collaboration on one side, suffering, persecution and resistance on the other. The United States, contending with internal and external violence, was trying to establish a multiracial society. At the same time it was fighting far from home for what, we were being told, was going to be a better world, though the reasons looked much less altruistic to our eyes. All these conflicts fostered our concerns. They also inspired the scientists of our generation to think about their role and responsibility as social beings. The nuclear era was a reminder that science was not as useless and abstruse as many had pretended it to be. Was it so desirable for us to be scientists during weekdays and citizens on Sundays and holidays, we asked ourselves. How could one justify indifference over educational matters, funding of universities, sexism, racism, and many other aspects of our daily existence? In thinking about a journal, questions like these were always present in our minds. COGNITION was born in France and we have edited the journal from its Paris office ever since. When Jacques Mehler moved from the United States to France, he worked in a laboratory located across from the Folies Bergeres, a neighborhood with many attractions for tourists but none of the scientific journals that were essential for keeping up with cognitive science. In 1969, the laboratory was moved to a modern building erected on the site at which the infamous Prison du Cherche-Midi had been located until its demolition at the end of the Second World War. This prison stood opposite the Gestapo Headquarters and resistance fighters and other personalities were tortured and then shot within its walls. A

Preface

IX

few decades earlier in the prison, another French citizen had been locked up, namely, Captain Dreyfus. It was difficult to find oneself at such a place without reflecting on how the rational study of the mind might illuminate the ways in which humans go about their social business and also, how science and society had to coexist. The building shelters the Ecole des Hautes Etudes en Sciences Sociales (EHESS), an institution that played an important role in the development of the French School of History. F. Braudel presided over the Ecole for many years while being the editor of the prestigious Annates, a publication that had won acclaim in many countries after it was founded by M. Bloch and L. Febvre. It was obvious that the Annates played an important role at the Ecole, where M. Bloch, an Alsatian Jew who was eventually murdered for his leading role during the Resistance, was remembered as an important thinker. Bloch was a convinced European who preached a rational approach to the social sciences. He was persuaded of the importance of expanding communication between investigators from different countries and cultures. Today, M. Bloch and his itinerary help us understand the importance of moral issues and the role of the individual as an ultimate moral entity whose well-being does not rank below state, country, or religion. Our hope is that rational inquiry and cognitive science will help us escape from the bonds of nationalism, chauvinism, and exclusion. Cognitive scientists, like all other scientists and citizens, should be guided by moral reason, and moral issues must be one of our fields of concern. A Dutch publisher, Mouton, offered us the opportunity to launch the journal. In the late sixties, money seemed less important than it does today. Publishers were interested in ideas and the elegance with which they were presented. We agreed to minimize formal constraints, and there was no opposition to the inclusion of a section to be used to air our political and social preoccupations. Opposition during those early planning stages came from a source that we had not at all foreseen as a trouble area. To our great surprise we discovered that publishing an English language journal in France was not an easy task. Some of our colleagues disapproved of what they perceived as a foreign-led venture. "Isn't it true," they argued, "that J. Piaget, one of the central players in the Cognitive Revolution, writes in French?" "A French intellectual ought to try and promote the French culture throughout the language of Descartes, Racine and Flaubert," we were reminded time and again. For a while we had mixed feelings. We need no reminders of how important differences and contrasts are to the richness of intellectual life. Today politicians discuss ways in which the world is going to be able to open markets and promote business. The GATT discussions have concentrated partly on the diversity of cultural goods. We agree with those who would like to see some kind of protection against mass-produced television, ghost-written books, and movies conceived to anesthetize the development of good taste and intelligence. Unfortunately, nobody really knows how to protect us against these lamentable trends. Removing all cultural differences and catering only to the least demanding members of society, no matter how numerous, will promote the destruction of our intellectual creativity. So why did we favor making a journal in English, and why is it that even today we fight for a lingua franca of science? Science is a special case, we told ourselves then, as we do today. We all know that since the Second World War, practically

X

Preface

all the top-quality science has been published in English. It would be unthinkable for top European scientists to have won the Nobel prize or reached world renown if they had published their foremost papers in their own language. They didn't. Likewise, it is unthinkable today for serious scientists, regardless of where they study and work, to be unable to read English. Of course, novels, essays, and many disciplines in the humanities are more concerned with form than with truth. It is normal that these disciplines fight to preserve their tool of privilege, the language in which they need to express themselves. Thus we viewed the resistance to English during the planning stages of COGNITION as an ideological plot to keep the study of mind separate and antagonistic to science and closer to the arts and humanities. Our aim was just the opposite, namely, to show that there was a discipline, cognition, which was as concerned with truth as chemistry, biology, or physics. We were also aware that the fear of contact and communication among fellow scientists is the favorite weapon used by narrow-minded chauvinists and, in general, by authoritarian characters with whom, we still, unfortunately, have to cope in some parts of the European academic world. While COGNITION was trying to impose the same weights and measures for European (inter alia) and American science, some of our colleagues were pleading for a private turf, for special journals catering to their specific needs. We dismissed those pleas, and the journal took the form that the readership has come to expect. Today, we include in this volume a series of articles that were originally published in the Special Issue produced to celebrate the fiftieth volume of the journal. We present these articles in an order which we think brings out their thematic coherence. There are areas that deal with theoretical aspects which range from the status of explanations in cognitive science, the evolutionary accounts offered to explain the stable faculties that are characteristic of homo abilis, to the way in which humans use general faculties to reason about their environment, and so forth. Another group of papers deals with the way in which humans process information and use language, the parts of cognitive science that are best understood, so far. We also present a number of papers that deal with infants' initial abilities and their capacity to learn the distinctive behaviors of the species. We also include several papers that try to relate behaviors to their underlying neural structures. This formto-function pairing may become particularly relevant to explain development. Indeed, many of the changes in behavior that one observes in the growing organism may stem from neural changes and/or from learning. Understanding the neural structures underlying our capacities may help us understand how these are mastered. It is difficult to imagine what the contents of volume 100 of COGNITION will look like. Certainly the journal, publishing in general, and academic publishing in particular, will change in radical ways in the years to come. Not only will the contents evolve in ways that will seem transparent a posteriori but also the form will change in ways that are hard to predict a priori. The ways in which science develops are hard to foresee because until one has bridged the next step vistas are occluded by the present. Fortunately, we do not need to worry about this for the time being. Our work is cut out—concentrating on what we are doing rather than on the ways in which we are doing what we are doing. On the

Preface

•

contrary, we must start thinking about how the changes in publishing will affect our ways of doing science. It is part of the scientist's duty to explore the changes to come so as to insure that the independence and responsibility of science is protected in the world of tomorrow as it is today. We cannot close this short introduction without thanking Amy Pierce for her help in preparing this special issue for publication with MIT Press. Jacques Mehler and Susana Franck

1 Insensitivity to future consequences following damage to human prefrontal cortex Antoine Bechara, Antonio R. Damasio*, Hanna Damasio, Steven W. Anderson Department of Neurology, Division of Behavioral Neurology and Cognitive Neuroscience, University of Iowa College of Medicine, Iowa City, IA 52242, USA

Abstract Following damage to the ventromedial prefrontal cortex, humans develop a defect in real-life decision-making, which contrasts with otherwise normal intellectual functions. Currently, there is no neuropsychological probe to detect in the laboratory, and the cognitive and neural mechanisms responsible for this defect have resisted explanation. Here, using a novel task which simulates real-life decision-making in the way it factors uncertainty of premises and outcomes, as well as reward and punishment, we find that prefrontal patients, unlike controls, are oblivious to the future consequences of their actions, and seem to be guided by immediate prospects only. This finding offers, for the first time, the possibility of detecting these patients' elusive impairment in the laboratory, measuring it, and investigating its possible causes.

Introduction Patients with damage to the ventromedial sector of prefrontal cortices develop a severe impairment in real-life decision-making, in spite of otherwise preserved intellect. The impairments are especially marked in the personal and social realms (Damasio, Tranel, & Damasio, 1991). Patient E.V.R. is a prototypical example of this condition. He often decides against his best interest, and is unable to learn

* Corresponding author. Supported by NINDS POl NS19632 and the James S. McDonnell Foundation.

4

A. Bechara, A. Damasio, H. Damasio, S. Anderson

from his mistakes. His decisions repeatedly lead to negative consequences. In striking contrast to this real-life decision-making impairment, E.V.R.'s general intellect and problem-solving abilities in a laboratory setting remain intact. For instance, he produces perfect scores on the Wisconsin Card Sorting Test (Milner, 1963), his performances in paradigms requiring self-ordering (Petrides & Milner, 1982), cognitive estimations (Shallice & Evans, 1978), and judgements of recency and frequency (Milner, Petrides, & Smith, 1985) are flawless; he is not preseverative, nor is he impulsive; his knowledge base is intact and so is his short-term and working memory as tested to date; his solution of verbally posed social problems and ethical dilemmas is comparable to that of controls (Saver & Damasio, 1991). The condition has posed a double challenge, since there has been neither a satisfactory account of its physiopathology, nor a laboratory probe to detect and measure an impairment that is so obvious in its ecological niche. Here we describe an experimental neuropsychological task which simulates, in real time, personal real-life decision-making relative to the way it factors uncertainty of premises and outcomes, as well as reward and punishment. We show that, unlike controls, patients with prefrontal damage perform defectively and are seemingly insensitive to the future.

Materials and methods The subjects sit in front of four decks of cards equal in appearance and size, and are given a $2000 loan of play money (a set of facsimile US bills). The subjects are told that the game requires a long series of card selections, one card at a time, from any of the four decks, until they are told to stop. After turning each card, the subjects receive some money (the amount is only announced after the turning, and varies with the deck). After turning some cards, the subjects are both given money and asked to pay a penalty (again the amount is only announced after the card is turned and varies with the deck and the position in the deck according to a schedule unknown to the subjects). The subjects are told that (1) the goal of the task is to maximize profit on the loan of play money, (2) they are free to switch from any deck to another, at any time, and as often as wished, but (3) they are not told ahead of time how many card selections must be made (the task is stopped after a series of 100 card selections). The preprogrammed schedules of reward and punishment are shown on the score cards (Fig. 1). Turning any card from deck A or deck B yields $100; turning any card from deck C or deck D yields $50. However, the ultimate future yield of each deck varies because the penalty amounts are higher in the high-paying decks (A and B), and lower in the low-paying decks (C and D). For example, after turning 10 cards from deck A, the subjects have earned $1000, but they have also encountered 5 unpredicted punishments bringing their total cost to $1250, thus

A TYPI

1

B 0

c

r

s

0

0 0 26 27 2 * 3 1

7

s

+50 5 1

r +50

3. 4

§

8

5

2/ 22 36 H9\ 50 si 52 53

2

+ 100

2

16

0

35 63

r

0

CM

CM

0

65 7tf7T

"£? 65 66

0

31

m CM

3i|37

Jo

0

7 ©

8 8 tf

0

72 73

38 39 4o

M2

wsM

55

56 59 6

of 0

0

0

0

CM]

A TYPICAL TARGET SUBJECT (DM 1336) 1 RESPONSE 1 OPTION

1 1

A

l 2

5

r

8

(/

1 B

/

1 +100

0

12 1

7

3 . 4 .5

7 .8

8 8 CM

8

I'M

9 10 H5 ^

o

1 2

u 58

g

8

0

0

3 59

4

5

60p

?

CM

22 31 32 33 55 57 65 66 * 7 68 69 1S\ CM

38 39 40

6

72 73 1H 7

8 0

CM

0

0

0

42 52 53 54 56 96 97 98 99 /oo

C

1 °

8 3 19 2

+5

/*

0

2 31 3

1 2

CM

5 16 J7

0

9 .10

23 2¥ 25 26 Z7 28 29 30 M3

8 *>

§

+100

+5

8

S

m

8 3

8

CM

B 49 56 51

D

1 ° 1.

9

6

7 10

0

77ie top score card repr represent the profiles of select more from decks

0

0

CM]

0

0

0

0

CM

0

that of a typical control subject, and the bottom one tha elections from the first to the 100th card. Control subjects B

6

A. Bechara, A. Damasio, H. Damasio, S. Anderson

incurring a net loss of $250. The same happens on deck B. On the other hand, after turning 10 cards from decks C or D, the subjects earn $500, but the total of their unpredicted punishments is only $250 (i.e. subject nets $250). In summary, decks A and B are equivalent in terms of overall net loss over the trials. The difference is that in deck A, the punishment is more frequent, but of smaller magnitude, whereas in deck B, the punishment is less frequent, but of higher magnitude. Decks C and D are also equivalent in terms of overall net loss. In deck C, the punishment is more frequent and of smaller magnitude, while in deck D the punishment is less frequent but of higher magnitude. Decks A and B are thus "disadvantageous" because they cost the most in the long run, while decks C and D are "advantageous" because they result in an overall gain in the long run. The performances of a group of normal control subjects (21 women and 23 men) in this task were compared to those of E.V.R. and other frontal lobe subjects (4 men and 2 women). The age range of normal controls was from 20 to 79 years; for E.V.R.-like subjects it was from 43 to 84 years. About half the number of subjects in each group had a high school education, and the other half had a college education. E.V.R.-like subjects were retrieved from the Patient Registry of the Division of Behavioral Neurology and Cognitive Neuroscience. Selection criteria were the documented presence of abnormal decision-making and the existence of lesions in the ventromedial prefrontal region. To determine whether the defective performance of E.V.R.-like subjects on the task is specific to ventromedial frontal lobe damage, and not merely caused by brain damage in general, we compared the performances of E.V.R.-like subjects and normal controls, to an education matched group of brain-damaged controls. There were 3 women and 6 men, ranging in age from 20 to 71 years. These controls were retrieved from the same Patient Registry and were chosen so as to have lesions in occipital, temporal and dorsolateral frontal regions. Several of the brain-damaged controls had memory defects, as revealed by conventional neuropsychological tests. Finally, to determine what would happen to the performance if it were repeated over time, we retested the target subjects and a smaller sample of normal controls (4 women and 1 man between the ages of 20 and 55, matched to E.V.R. in level of education) after various time intervals (one month after the first test, 24 h later, and for the fourth time, six months later).

Results Fig. 2 (left) shows that normal controls make more selections from the good decks (C and D), and avoid the bad decks (A and B). In sharp contrast, E.V.R.-like subjects select fewer from the good decks (C and D), and choose more from the bad decks (A and B). The difference is significant. An analysis of

A TYPICAL CON

NORMAL CONTROLS

A

i SJ f

HIIIIIIIIIIII

f 1111111111111111 i 11111111 f 11111111111111 • 111111111111

0

15

30

45

A TYPICAL TARGET SUBJ

IATL D B

C

DISADVANTAGEOUS

HHIIIH

t

liiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiiii 0 15 30 45

ADVANTAGEOUS

Choice Fig. 2.

^J

r Tria

(Left panels) Total number of cards selected from each deck (A, B, C or D) by normal controls (n = represent means ± s.e.m. (Right panels) Profiles of card selections (from the first to the 100th selection)

A. Bechara, A. Damasio, H. Damasio, S. Anderson

8

variance comparing the number of cards from each deck chosen by normal controls and by target subjects revealed a significant interaction of group (controls vs. targets) with choice (A, B, C, D) (F(3,147) = 42.9, /X.001). Subsequent Newman-Keuls Mests revealed that the number of cards selected by normal controls from deck A or B were significantly less than the number of cards selected by target subjects from the same decks (ps< .001). On the contrary, the number of cards selected by controls from decks C or D were significantly higher than the numbers selected by target subjects (ps

60-1

S

4 WD

ac« >• •o < «M O u

1Za

Fig. 3.

-20/f A

-40-

Normal

Brain- Damaged

CONTROLS

EVR

EVR-Typc

TARGET SUBJECTS

Total number of selections from the advantageous decks (C + D) minus the total numbers of selections from the disadvantageous decks (A + B) from a group of normal controls (n = 44), brain-damaged controls (n=9), E.V.R., and E.V.R.-like subjects (n=6). Bars represent means ± s.e.m. Positive scores reflect advantageous courses of action, and negative scores reflect disadvantageous courses of action.

Insensitivity to future consequences following damage to prefrontal cortex

9

from that of normal controls, and quite the opposite of the performance of the prefrontal subjects. One-way ANOVA on the difference in the total numbers of card selections from the advantageous decks minus the total numbers of selections from the disadvantageous decks obtained from normal and brain-damaged controls did not reveal a significant difference between the two groups (F(l,52) = 0.1, p> A), but the difference between the normal and E.V.R.-like groups was highly significant (F(l,50) = 74.8, p < .001). As a result of repeated testing, E.VR.'s performance did not change, one way or the other, when tested one month after the first test, 24 h later, and for the fourth time, six months later. This pattern of impaired performance was also seen in other target subjects. On the contrary, the performance of normal controls improved over time.

Discussion These results demonstrate that E.V.R. and comparable subjects perform defectively in this task, and that the defect is stable over time. Although the task involves a long series of gains and losses, it is not possible for subjects to perform an exact calculation of the net gains or losses generated from each deck as they play. Indeed, a group of normal control subjects with superior memory and IQ, whom we asked to think aloud while performing the task, and keep track of the magnitudes and frequencies of the various punishments, could not provide calculated figures of the net gains or losses from each deck. The subjects must rely on their ability to develop an estimate of which decks are risky and which are profitable in the long run. Thus, the patients' performance profile is comparable to their real-life inability to decide advantageously, especially in personal and social matters, a domain for which in life, as in the task, an exact calculation of the future outcomes is not possible and choices must be based on approximations. We believe this task offers, for the first time, the possibility of detecting these patients' elusive impairment in the laboratory, measuring it, and investigating its possible causes. Why do E.V.R.-like subjects make choices that have high immediate reward, but severe delayed punishment? We considered three possibilities: (1) E.V.R.-like subjects are so sensitive to reward that the prospect of future (delayed) punishment is outweighed by that of immediate gain; (2) these subjects are insensitive to punishment, and thus the prospect of reward always prevails, even if they are not abnormally sensitive to reward; (3) these subjects are generally insensitive to future consequences, positive or negative, and thus their behavior is always guided by immediate prospects, whatever they may be. To decide on the merit of these possibilities, we developed a variant of the basic task, in which the schedules of reward and punishment were reversed, so that the punishment is immediate and

10

A. Bechara, A. Damasio, H. Damasio, S. Anderson

the reward is delayed. The profiles of target subjects in that task suggest that they were influenced more by immediate punishment than by delayed reward (unpublished results). This indicates that neither insensitivity to punishment nor hypersensitivity to reward are appropriate accounts for the defect. A qualitative aspect of the patients' performance also supports the idea that immediate consequences influence the performance significantly. When they are faced with a significant money loss in a given deck, they refrain from picking cards out of that same deck, for a while, just like normals do, though unlike normals they then return to select from that deck after a few additional selections. When we combine the profiles of both basic task and variant tasks, we are left with one reasonable possibility: that these subjects are unresponsive to future consequences, whatever they are, and are thus more controlled by immediate prospects. How can this "myopia" for the future be explained? Evidence from other studies suggests that these patients possess and can access the requisite knowledge to conjure up options of actions and scenarios of future outcomes just as normal controls do (Saver & Damasio, 1991). Their defect seems to be at the level of acting on such knowledge. There are several plausible accounts to explain such a defect. For instance, it is possible that the representations of future outcomes that these patients evoke are unstable, that is, that they are not held in working memory long enough for attention to enhance them and reasoning strategies to be applied to them. This account invokes a defect along the lines proposed for behavioral domains dependent on dorsolateral prefrontal cortex networks, and which is possibly just as valid in the personal/social domain of decision-making (Goldman-Rakic, 1987). Defects in temporal integration and attention would fall under this account (Fuster, 1989; Posner, 1986). Alternatively, the representations of future outcomes might be stable, but they would not be marked with a negative or positive value, and thus could not be easily rejected or accepted. This account invokes the somatic marker hypothesis which posits that the overt or covert processing of somatic states provides the value mark for a cognitive scenario (Damasio, 1994; Damasio et al., 1991). We have been attempting to distinguish between these two accounts in a series of subsequent experiments using this task along with psychophysiological measurements. Preliminary results favor the latter account, or a combination of the two accounts. Those results also suggest that the biasing effect of the value mark operates covertly, at least in the early stages of the task.

References Damasio, A.R. (1994). Descartes' error: Emotion, rationality and the human brain. New York: Putnam (Grosset Books).

Insensitivity to future consequences following damage to prefrontal cortex

11

Damasio, A.R., Tranel, D., & Damasio, H. (1991). Somatic markers and the guidance of behavior. In H. Levin, H. Eisenberg, & A. Benton (Eds.), Frontal lobe function and dysfunction (pp. 217-228). New York: Oxford University Press. Fuster, J.M. (1989). The prefrontal cortex (2nd edn.). New York: Raven Press. Goldman-Rakic, P.S. (1987). Circuitry of primate prefrontal cortex and regulation of behavior by representational memory. In F. Plum (Ed.), Handbook of physiology: The nervous system (Vol. V, pp. 373-401). Bethesda, MD: American Physiological Society. Milner, B. (1963). Effects of different brain lesions on card sorting. Archives of Neurology, 9, 90-100. Milner! B., Petrides, M., & Smith, M.L. (1985). Frontal lobes and the temporal organization of memory. Human Neurobiology, 4, 137-142. Petrides, M., & Milner, B (1982). Deficits on subject-ordered tasks after frontal and temporal-lobe lesions in man. Neuropsychologia, 20, 249-262. Posner, M.I. (1986). Chronometric explorations of the mind. New York: Oxford University Press. Saver, J.L., & Damasio, A.R. (1991). Preserved access and processing of social knowledge in a patient with acquired sociopathy due to ventromedial frontal damage. Neuropsychologia, 29, 1241-1249. Shallice, T., & Evans, M.E. (1978). The involvement of the frontal lobes in cognitive estimation. Cortex, 14, 294-303.

2

Autism: beyond "theory of mind" Uta Frith*, Francesca Happe MRC Cognitive Development Unit, 4 Taviton Street, London WC1H OBT, UK

Abstract The theory of mind account of autism has been remarkably successful in making specific predictions about the impairments in socialization, imagination and communication shown by people with autism. It cannot, however, explain either the non-triad features of autism, or earlier experimental findings of abnormal assets and deficits on non-social tasks. These unexplained aspects of autism, and the existence of autistic individuals who consistently pass false belief tasks, suggest that it may be necessary to postulate an additional cognitive abnormality. One possible abnormality - weak central coherence - is discussed, and preliminary evidence for this theory is presented.

The theory of mind account of autism In 1985 Cognition published an article by Baron-Cohen, Leslie, and Frith, entitled: Does the autistic child have a "theory of mind"? The perceptive reader would have recognized this as a reference to Premack and Woodruffs (1978) question: Does the chimpanzee have a theory of mind? The connection between these two was, however, an indirect one - the immediate precursor of the paper was Wimmer and Perner's (1983) article on the understanding of false beliefs by normally developing pre-school children. Each of these three papers has, in its way, triggered an explosion of research interest; in the social impairments of autism, the mind-reading capacities of non-human primates, and the development of social understanding in normal children. The connections which existed between the three papers have been mirrored in continuing connections between these three fields of research - developmental psychology (Astington, Harris, & Olson, 1989; Perner, 1991; Russell, 1992; Wellman, 1990), cognitive ethology

* Corresponding author

14

U. Frith, F. Happ£

(Byrne & Whiten, 1988; Cheney & Seyfarth, 1990), and developmental psychopathology (Cicchetti & Cohen, in press; Rutter, 1987). There can be little doubt that these contacts have enriched work in each area. Perceptive readers would also have noticed the inverted commas surrounding the phrase "theory of mind" in the 1985 paper. Baron-Cohen, Leslie, and Frith followed Premack and Woodruffs definition of this "sexy" but misleading phrase: to have a theory of mind is to be able to attribute independent mental states to self and others in order to explain and predict behaviour. As might befit a "theory" ascribable to chimpanzees, this was not a conscious theory but an innately given cognitive mechanism allowing a special sort of representation - the representation of mental states. Leslie (1987, 1988) delivered the critical connection between social understanding and understanding of pretence, via this postulated mechanism; metarepresentation is necessary, in Leslie's theory, for representing pretence, belief and other mental states. From this connection, between the social world and the world of imaginative play, sprung the link to autistic children, who are markedly deficient in both areas. The idea that people with autism could be characterized as suffering from a type of "mind-blindness", or lack of theory of mind, has been useful to the study of child development - not because it was correct (that is still debatable) but because it was a causal account which was both specific and falsifiable. The clearest expression of this causal account is given in Frith, Morton, and Leslie (1991). What is to be explained? Autism is currently defined at the behavioural level, on the basis of impairments in socialization, communication and imagination, with stereotyped repetitive interests taking the place of creative play (DSM-III-R, American Psychological Association, 1987). A causal account must link these behavioural symptoms to the presumed biological origins (Gillberg & Coleman, 1992; Schopler & Mesibov, 1987) of this disorder. Specificity is particularly important in any causal account of autism because autistic people themselves show a highly specific pattern of deficits and skills. The IQ profile alone serves to demonstrate this; autistic people in general show an unusually "spiky" profile across Wechsler subtests (Lockyer & Rutter, 1970; Tymchuk, Simmons, & Neafsey, 1977), excelling on Block Design (constructing a pattern with cubes), and failing on Picture Arrangement (ordering pictures in a cartoon strip). This puzzling discrepancy of functioning has caused many previous psychological theories of autism to fail. For example, high arousal, lack of motivation, language impairment, or perceptual problems are all too global to allow for both the assets and deficits of autism.

Fine cuts along a hidden seam What are the specific predictions made by the hypothesis that people with autism lack a "theory of mind"? The hypothesis does not address the question of

15

Autism: beyond "theory of mind"

the spiky IQ profile - it is silent on functioning in non-social areas - but it focuses on the critical triad of impairments (Wing & Gould, 1979). Not only does it make sense of this triad, but it also makes "fine cuts" within the triad of autistic impairments. Social and communicative behaviour is not all of one piece, when viewed from the cognitive level. Some, but not all, such behaviour requires the ability to "mentalize" (represent mental states). So, for example, social approach need not be built upon an understanding of others' thoughts - indeed Hermelin and O'Connor (1970) demonstrated to many people's initial surprise that autistic children prefer to be with other people, just like non-autistic children of the same mental age. However, sharing attention with someone else does require mentalizing - and is consistently reported by parents to be missing in the development of even able autistic children (Newson, Dawson, & Everard, 1984). The mentalizing-deficit account has allowed a systematic approach to the impaired and unimpaired social and communicative behaviour of people with autism. Table 1 shows some of the work exploring predictions from the hypothesis that autistic people lack mentalizing ability. The power of this hypothesis is to make fine cuts in the smooth continuum of behaviours, and in this it has been remarkably useful. It has sparked an enormous amount of research, both supporting and attacking the theory (reviewed by Baron-Cohen, Tager-Flusberg, & Cohen, 1993; Happe, 1994a; Happe & Frith, in press). The fine cuts method, as used in the laboratory, has also informed research

Table 1. Autistic assets and deficits as predicted by the "fine cuts" technique, between tasks which require mentalizing and those which do not Assets

Deficits

Ordering behavioural pictures

Ordering mentalistic pictures (Baron-Cohen et al., 1986) Understanding know (Perner et al., 1989) Protodeclarative pointing (Baron-Cohen, 1989b) Deception (Sodian & Frith, 1992) False beliefs (Leslie & Thaiss, 1992; Leekam & Perner, 1991) Recognizing surprise (Baron-Cohen et al., 1993) Information occlusion (Baron-Cohen, 1992) Metaphorical expression (Happe, 1993)

Understanding see Protoimperative pointing Sabotage False photographs

Recognizing happiness and sadness Object occlusion Literal expression References refer to Assets and Deficits.

16

U. Frith, F. Happe*

Table 2. Autistic assets and deficits observed in real life Assets

Deficits

Elicited structured play

Spontaneous pretend play (Wetherby & Prutting, 1984) Expressive gestures (Attwood, Frith, & Hermelin, 1988) Talking about beliefs and ideas (Tager-Flusberg, 1993) Using person as receiver of information (Phillips, 1993) Showing "interactive" sociability (Frith et al., in press)

Instrumental gestures Talking about desires and emotions Using person as tool Showing "active" sociability ^ References refer to Assets and Deficits.

into the pattern of abilities and deficits in real life (Table 2), although this enterprise has still some way to go. This technique, which aims to pit two behaviours against each other which differ only in the demands they make upon the ability to mentalize, pre-empts many potential criticisms. It is also peculiarly suitable for use in brain-imaging studies. By looking at performance across tasks which are equivalent in every other way, except for the critical cognitive component, intellectual energy has been saved for the really interesting theoretical debates. Another key benefit of the specificity of this approach is the relevance it has for normal development. The fine cuts approach suits the current climate of increased interest in the modular nature of mental capacities (e.g., Cosmides, 1989; Fodor, 1983). It has allowed us to think about social and communicative behaviour in a new way. For this reason, autism has come to be a test case for many theories of normal development (e.g., Happe, 1993; Sperber & Wilson's 1986 Relevance theory).

Limitations of the theory of mind account The hijacking of autism by those primarily interested in normal development has added greatly to the intellectual richness of autism research. But just how well does the theory of mind account explain autism? By the stringent standard, that explanatory theories must give a full account of a disorder (Morton & Frith, in press), not that well. The mentalizing account has helped us to understand the nature of the autistic child's impairments in play, social interaction and verbal and non-verbal communication. But there is more to autism than the classic triad of impairments.

Autism: beyond "theory of mind"

17

Non-triad features Clinical impressions originating with Kanner (1943) and Asperger (1944; translated in Frith, 1991), and withstanding the test of time, include the following: - Restricted repertoire of interests (necessary for diagnosis in DSM-III-R, American Psychological Association, 1987). - Obsessive desire for sameness (one of two cardinal features for Kanner & Eisenberg, 1956). - Islets of ability (an essential criterion in Kanner, 1943). - Idiot savant abilities (striking in 1 in 10 autistic children, Rimland & Hill, 1984). - Excellent rote memory (emphasized by Kanner, 1943). - Preoccupation with parts of objects (a diagnostic feature in DSM-IV, forthcoming). All of these non-triad aspects of autism are vividly documented in the many parental accounts of the development of autistic children (Hart, 1989; McDonnell, 1993; Park, 1967). None of these aspects can be well explained by a lack of mentalizing. Of course, clinically striking features shown by people with autism need not be specific features of the disorder. However, there is also a substantial body of experimental work, much of it predating the mentalizing theory, which demonstrates non-social abnormalities that are specific to autism. Hermelin and O'Connor were the first to introduce what was in effect a different "fine cuts" method (summarized in their 1970 monograph) - namely the comparison of closely matched groups of autistic and non-autistic handicapped children of the same mental age. Table 3 summarizes some of the relevant findings.

The talented minority The mentalizing deficit theory of autism, then, cannot explain all features of autism. It also cannot explain all people with autism. Even in the first test of the hypothesis (reported in the 1985 Cognition paper), some 20% of autistic children passed the Sally-Ann task. Most of these successful children also passed another test of mentalizing - ordering picture stories involving mental states (BaronCohen, Leslie, & Frith, 1986) - suggesting some real underlying competence in representing mental states. Baron-Cohen (1989a) tackled this apparent dis-

18

U. Frith, F. Happe*

Table 3. Experimental findings not accounted for by mind-blindness. Surprising advantages and disadvantages on cognitive tasks, shown by autistic subjects relative to normally expected asymmetries Unusual strength

Unusual weakness

Memory for word strings

Memory for sentences (e.g., Hermelin & O'Connor, 1967) Memory for related items (e.g., Tager-Flusberg, 1991) Echoing with repair (e.g., Aurnhammer-Frith, 1969) Pattern detection (e.g., Frith, 1970 a,b) Jigsaw by picture (e.g., Frith & Hermelin, 1969) Sorting faces by person (e.g., Weeks & Hobson, 1987) Recognizing faces right-way-up (e.g., Langdell, 1978)

Memory for unrelated items Echoing nonsense Pattern imposition Jigsaw by shape Sorting faces by accessories Recognizing faces upside-down

References refer to Unusual strength and Unusual weakness.

confirmation of the theory, by showing that these talented children still did not pass a harder (second-order) theory of mind task (Perner & Wimmer, 1985). However, results from other studies focusing on high-functioning autistic subjects (Bowler, 1992; Ozonoff, Rogers, & Pennington, 1991) have shown that some autistic people can pass theory of mind tasks consistently, applying these skills across domains (Happe, 1993) and showing evidence of insightful social behaviour in everyday life (Frith, Happe, & Siddons, in press). One possible way of explaining the persisting autism of these successful subjects is to postulate an additional and continuing cognitive impairment. What could this impairment be? The recent interest in executive function deficits in autism (Hughes & Russell, 1993; Ozonoff, Pennington, & Rogers, 1991) can be seen as springing from some of the limitations of the theory of mind view discussed above. Ozonoff, Rogers, & Pennington (1991) found that while not all subjects with autism and /or Asperger's syndrome showed a theory of mind deficit, all were impaired on the Wisconsin Card Sorting Test and Tower of Hanoi (two typical tests of executive function). On the basis of this finding they suggest that executive function impairments are a primary causal factor in autism. However, the specificity, and hence the power of this theory as a causal account, has yet to be established by systematic comparison with other non-autistic groups who show impairments in executive functions (Bishop, 1993). While an additional impairment in executive functions may be able to explain certain (perhaps non-specific) features of autism (e.g., stereotypies, failure to plan, impulsiveness), it is not clear how it could explain the specific deficits and skills summarized in Table 3.

Autism: beyond "theory of mind"

19

The central coherence theory Motivated by the strong belief that both the assets and the deficits of autism spring from a single cause at the cognitive level, Frith (1989) proposed that autism is characterized by a specific imbalance in integration of information at different levels. A characteristic of normal information processing appears to be the tendency to draw together diverse information to construct higher-level meaning in context; "central coherence" in Frith's words. For example, the gist of a story is easily recalled, while the actual surface form is quickly lost, and is effortful to retain. Bartlett (1932), summarizing his famous series of experiments on remembering images and stories, concluded: "an individual does not normally take [such] a situation detail by detail... In all ordinary instances he has an overmastering tendency simply to get a general impression of the whole; and, on the basis of this, he constructs the probable detail" (p. 206). Another instance of central coherence is the ease with which we recognize the contextually appropriate sense of the many ambiguous words used in everyday speech (son-sun, meet-meat, sew-so, pear-pair). A similar tendency to process information in context for global meaning is also seen with non-verbal material - for example, our everyday tendency to misinterpret details in a jigsaw piece according to the expected position in the whole picture. It is likely that this preference for higher levels of meaning may characterize even mentally handicapped (non-autistic) individuals - who appear to be sensitive to the advantage of recalling organized versus jumbled material (e.g., Hermelin & O'Connor, 1967). Frith suggested that this universal feature of human information processing was disturbed in autism, and that a lack of central coherence could explain very parsimoniously the assets and deficits shown in Table 3. On the basis of this theory, she predicted that autistic subjects would be relatively good at tasks where attention to local information - relatively piece-meal processing - is advantageous, but poor at tasks requiring the recognition of global meaning.

Empirical evidence: assets A first striking signpost towards the theory appeared quite unexpectedly, when Amitta Shah set off to look at autistic children's putative perceptual impairments on the Embedded Figures Test. The children were almost better than the experimenter! Twenty autistic subjects with an average age of 13, and non-verbal mental age of 9.6, were compared with 20 learning disabled children of the same age and mental age, and 20 normal 9-year-olds. These children were given the Children's Embedded Figures Test (CEFT; Witkin, Oltman, Raskin, & Karp, 1971), with a slightly modified procedure including some pretraining with cut-out shapes. The test involved spotting a hidden figure (triangle or house shape)

20

U. Frith, F. Happ6

among a larger meaningful drawing (e.g., a clock). During testing children were allowed to indicate the hidden figure either by pointing or by using a cut-out shape of the hidden figure. Out of a maximum score of 25, autistic children got a mean of 21 items correct, while the two control groups (which did not differ significantly in their scores) achieved 15 or less. Gottschaldt (1926) ascribed the difficulty of finding embedded figures to the overwhelming "predominance of the whole". The ease and speed with which autistic subjects picked out the hidden figure in Shah and Frith's (1983) study was reminiscent of their rapid style of locating tiny objects (e.g. thread on a patterned carpet) and their immediate discovery of minute changes in familiar lay-outs (e.g., arrangement of cleaning materials on bathroom shelf), as often described anecdotally. The study of embedded figures was introduced into experimental psychology by the Gestalt psychologists, who believed that an effort was needed to resist the tendency to see the forcefully created gestalt, at the expense of the constituent parts (Koffka, 1935). Perhaps this struggle to resist overall gestalt forces does not occur for autistic subjects. If people with autism, due to weak central coherence, have privileged access to the parts and details normally securely embedded in whole figures, then novel predictions could be made about the nature of their islets of ability. The Block Design subtest of the Wechsler Intelligence Scales (Wechsler, 1974, 1981) is consistently found to be a test on which autistic people show superior performance relative to other subtests, and often relative to other people of the same age. This test, first introduced by Kohs (1923), requires the breaking up of line drawings into logical units, so that individual blocks can be used to reconstruct the original design from separate parts. The designs are notable for their strong gestalt qualities, and the difficulty which most people experience with this task appears to relate to problems in breaking up the whole design into the constituent blocks. While many authors have recognized this subtest as an islet of ability in autism, this fact has generally been explained as due to intact or superior general spatial skills (Lockyer & Rutter, 1970; Prior, 1979). Shah and Frith (1993) suggested, on the basis of the central coherence theory, that the advantage shown by autistic subjects is due specifically to their ability to see parts over wholes. They predicted that normal, but not autistic, subjects would benefit from pre-segmentation of the designs. Twenty autistic, 33 normal and 12 learning disabled subjects took part in an experiment, where 40 different block designs had to be constructed from either whole or pre-segmented drawn models (Fig. 1). Autistic subjects with normal or near-normal non-verbal IQ were matched with normal children of 16 years. Autistic subjects with non-verbal IQ below 85 (and not lower than 57) were compared with learning disabled children of comparable IQ and chronological age (18 years), and normal children aged 10. The results showed that the autistic subjects' skill on this task resulted from a greater ability to segment the design. Autistic subjects showed superior performance compared to controls in one

21

Autism: beyond "theory of mind"

HH

• «• H

BB

3

4> 8

Fig. 1. Examples of all types of design: "whole" versus "segmented" (1, 2, 3, 4 vs. 5, 6, 7, 8) "oblique" versus "non-oblique" (3, 4, 7, 8 vs. 1, 2, 5, 6) "unrotated" versus "rotated" (1, 3, 5, 7 vs. 2, 4, 6, 8).

condition only - when working from whole designs. The great advantage which the control subjects gained from using pre-segmented designs was significantly diminished in the autistic subjects, regardless of their IQ level. On the other hand, other conditions which contrasted presence and absence of obliques, and rotated versus unrotated presentation, affected all groups equally. From these latter findings it can be concluded that general visuo-spatial factors show perfectly normal effects in autistic subjects, and that superior general spatial skill may not account for Block design superiority.

Empirical evidence: deficits While weak central coherence confers significant advantages in tasks where preferential processing of parts over wholes is useful, it would be expected to confer marked disadvantages in tasks which involve interpretation of individual

22

U. Frith, F. Happ6

stimuli in terms of overall context and meaning. An interesting example is the processing of faces, which seems to involve both featural and configural processing (Tanka & Farah, 1993). Of these two types of information, it appears to be configural processing which is disrupted by the inverted presentation of faces (Bartlett & Searcy, 1993; Rhodes, Brake, & Atkinson, 1993). This may explain the previously puzzling finding that autistic subjects show a diminished disadvantage in processing inverted faces (Hobson, Ouston, & Lee, 1988; Langdell, 1978). One case in which the meaning of individual stimuli is changed by their context is in the disambiguation of homographs. In order to choose the correct (contextappropriate) pronunciation in the following sentences, one must process the final word as part of the whole sentence meaning: "He had a pink bow"; "He made a deep bow". Frith and Snowling (1983) predicted that this sort of contextual disambiguation would be problematic for people with autism. They tested 8 children with autism who had reading ages of 8-10 years, and compared them with 6 dyslexic children and 10 normal children of the same reading age. The number of words read with the contextually appropriate pronunciation ranged from 5 to 7 out of 10 for the autistic children, who tended to give the more frequent pronunciation regardless of sentence context. By contrast, the normal and dyslexic children read between 7 and 9 of the 10 homographs in a contextually determined manner. This finding suggested that autistic children, although excellent at decoding single words, were impaired when contextual cues had to be used. This was also demonstrated in their relative inability to answer comprehension questions and to fill in gaps in a story text. This work fits well with previous findings (Table 3) concerning failure to use meaning and redundancy in memory tasks.

The abnormality of excellence The hypothesis that people with autism show weak central coherence aims to explain both the glaring impairments and the outstanding skills of autism as resulting from a single characteristic of information processing. One characteristic of this theory is that it claims that the islets of ability and savant skills are achieved through relatively abnormal processing, and predicts that this may be revealed in abnormal error patterns. One example might be the type of error made in the Block Design test. The central coherence theory suggests that, where errors are made at all on Block Design, these will be errors which violate the overall pattern, rather than the details. Kramer, Kaplan, Blusewicz, and Preston (1991) found that in normal adult subjects there was a strong relation between the number of such configuration-breaking errors made on the Block Design test and the number of local (vs. global) choices made in a similarity-judgement task

Autism: beyond "theory of mind"

23

(Kimchi & Palmer, 1982). Preliminary data from subjects with autism (Happe, in preparation) suggest that, in contrast to normal children, errors violating configuration are far more common than errors violating pattern details in autistic Block Design performance. A second example concerns idiot savant drawing ability. Excellent drawing ability may be characterized by a relatively piece-meal drawing style. Mottron and Belleville (1993) found in a case study of one autistic man with exceptional artistic ability that performance on three different types of tasks suggested an anomaly in the hierarchical organization of the local and global parts of figures. The authors observed that the subject "began his drawing by a secondary detail and then progressed by adding contiguous elements", and concluded that his drawings showed "no privileged status of the global form . . . but rather a construction by local progression". In contrast, a professional draughtsman who acted as a control started by constructing outlines and then proceeded to parts. It remains to be seen whether other savant abilities can be explained in terms of a similarly local and detail-observant processing style.

Central coherence and mentalizing Central coherence, then, may be helpful in explaining some of the real-life features that have so far resisted explanation, as well as making sense of a body of experimental work not well accounted for by the mentalizing deficit theory. Can it also shed light on the continuing handicaps of those talented autistic subjects who show consistent evidence of some mentalizing ability? Happe (1991), in a first exploration of the links between central coherence and theory of mind, used Snowling and Frith's (1986) homograph reading task with a group of able autistic subjects. Autistic subjects were tested on a battery of theory of mind tasks at two levels of difficulty (first- and second-order theory of mind), and grouped according to their performance (Happe, 1993). Five subjects who failed all the theory of mind tasks, 5 subjects who passed all and only first-order tasks, and 6 subjects who passed both first- and second-order theory of mind tasks were compared with 14 7-8-year-olds. The autistic subjects were of mean age 18 years, and had a mean IQ of around 80. The three autistic groups and the control group obtained the same score for total number of words correctly read. As predicted, however, the young normal subjects, but not the autistic subjects, were sensitive to the relative position of target homograph and disambiguating context: "There was a big tear in her eye", versus "In her dress there was a big tear". The normal controls showed a significant advantage when sentence context occurred before (rare pronunciation) target words (scoring 5 out of 5, vs. 2 out of 5 where target came first), while the autistic subjects (as in Frith and Snowling, 1983) tended to give the more frequent pronunciation regardless (3 out of 5 appropriate pronun-

24

U. Frith, F. Happ6

ciations in each case). The important point of this study was that this was true of all three autistic groups, irrespective of level of theory of mind performance. Even those subjects who consistently passed all the theory of mind tasks (mean VIQ 90) failed to use sentence context to disambiguate homograph pronunciation. It is possible, therefore, to think of weak central coherence as characteristic of even those autistic subjects who possess some mentalizing ability. Happe (submitted) explored this idea further by looking at WISC-R and WAIS subtest profiles. Twenty-seven children who failed standard first-order false belief tasks were compared with 21 subjects who passed. In both groups Block Design was a peak of non-verbal performance for the majority of subjects: 18/21 passers, and 23/27 failers. In contrast, performance on the Comprehension subtest (commonly thought of as requiring pragmatic and social skill) was a low point in verbal performance for 13/17 "failers" but only 6/20 "passers". It seems, then, that while social reasoning difficulties (as shown by Wechsler tests) are striking only in those subjects who fail theory of mind tasks, skill on non-verbal tasks benefiting from weak central coherence is characteristic of both passers and failers. There is, then, preliminary evidence to suggest that the central coherence hypothesis is a good candidate for explaining the persisting handicaps of the talented minority. So, for example, when theory of mind tasks were embedded in slightly more naturalistic tasks, involving extracting information from a story context, even autistic subjects who passed standard second-order false belief tasks showed characteristic and striking errors of mental state attribution (Happe, 1994b). It may be that a theory of mind mechanism which is not fed by rich and integrated contextual information is of little use in everyday life. The finding that weak central coherence may characterize autistic people at all levels of theory of mind ability goes against Frith's (1989) original suggestion that a weakness in central coherence could by itself account for theory of mind impairment. At present, all the evidence suggests that we should retain the idea of a modular and specific mentalizing deficit in our causal explanation of the triad of impairments in autism. It is still our belief that nothing captures the essence of autism so precisely as the idea of "mind-blindness". Nevertheless, for a full understanding of autism in all its forms, this explanation alone will not suffice. Therefore, our present conception is that there may be two rather different cognitive characteristics that underlie autism. Following Leslie (1987, 1988) we hold that the mentalizing deficit can be usefully conceptualized as the impairment of a single modular system. This system has a neurological basis - which may be damaged, leaving other functions intact (e.g., normal IQ). The ability to mentalize would appear to be of such evolutionary value (Byrne & Whiten, 1988; Whiten, 1991) that only insult to the brain can produce deficits in this area. By contrast, the processing characteristic of weak central coherence, as illustrated above, gives both advantages and disadvantages, as would strong central coherence. It is possible, then, to think of this balance (between preference for parts

Autism: beyond "theory of mind"

25

vs. wholes) as akin to a cognitive style, which may vary in the normal population. No doubt, this style would be subject to environmental influences, but, in addition, it may have a genetic component. It may be interesting, then, to focus on the strengths and weaknesses of autistic children's processing, in terms of weak central coherence, in looking for the extended phenotype of autism. Some initial evidence for this may be found in the report by Landa, Folstein, and Isaacs (1991) that the parents of children with autism tell rather less coherent spontaneous narratives than do controls.

Central coherence and executive function With the speculative link to cognitive style rather than straightforward deficit, the central coherence hypothesis differs radically not only from the theory of mind account, but also from other recent theories of autism. In fact, every other current psychological theory claims that some significant and objectively harmful deficit is primary in autism. Perhaps the most influential of such general theories is the idea that autistic people have executive function deficits, which in turn cause social and non-social abnormalities. The umbrella term "executive functions" covers a multitude of higher cognitive functions, and so is likely to overlap to some degree with conceptions of both central coherence and theory of mind. However, the hypothesis that autistic people have relatively weak central coherence makes specific and distinct predictions even within the area of executive function. For example, the "inhibition of pre-potent but incorrect responses" may contain two separable elements: inhibition and recognition of context-appropriate response. One factor which can make a pre-potent response incorrect is a change of context. If a stimulus is treated in the same way regardless of context, this may look like a failure of inhibition. However, autistic people may have no problem in inhibiting action where context is irrelevant. Of course it may be that some people with autism do have an additional impairment in inhibitory control, just as some have peripheral perceptual handicaps or specific language problems.

Future prospects The central coherence account of autism is clearly still tentative and suffers from a certain degree of over-extension. It is not clear where the limits of this theory should be drawn - it is perhaps in danger of trying to take on the whole problem of meaning! One of the areas for future definition will be the level at which coherence is weak in autism. While Block Design and Embedded Figures tests appear to tap processing characteristics at a fairly low or perceptual level,

26

U. Frith, F. Happ*

work on memory and verbal comprehension suggests higher-level coherence deficits. Coherence can be seen at many levels in normal subjects, from the global precedence effect in perception of hierarchical figures (Navon, 1977) to the synthesis of large amounts of information and extraction of inferences in narrative processing (e.g., Trabasso & Suh, 1993, in a special issue of Discourse Processes on inference generation during text comprehension). One interesting way forward may be to contrast local coherence within modular systems, and global coherence across these systems in central processing. So, for example, the calendrical calculating skills of some people with autism clearly show that information within a restricted domain can be integrated and processed together (O'Connor & Hermelin, 1984; Hermelin & O'Connor, 1986), but the failure of many such savants to apply their numerical skills more widely (some cannot multiply two given numbers) suggests a modular system specialized for a very narrow cognitive task. Similarly, Norris (1990) found that building a connectionist model of an "idiot savant date calculator" only succeeded when forced to take a modular approach. Level of coherence may be relative. So, for example, within text there is the word-to-word effect of local association, the effect of sentence context, and the larger effect of story structure. These three levels may be dissociable, and it may be that people with autism process the most local of the levels available in open-ended tasks. The importance of testing central coherence with open-ended tasks is suggested by a number of findings. For example, Snowling and Frith (1986) demonstrated that it was possible to train subjects with autism to give the context appropriate (but less frequent) pronunciation of ambiguous homographs. Weeks and Hobson (1987) found that autistic subjects sorted photographs of faces by type of hat when given a free choice, but, when asked again, were able to sort by facial expression. It seems likely, then, that autistic weak central coherence is most clearly shown in (non-conscious) processing preference, which may reflect the relative cost of two types of processing (relatively global and meaningful vs. relatively local and piece-meal). Just as the idea of a deficit in theory of mind has taken several years and considerable (and continuing) work to be empirically established, so the idea of a weakness in central coherence will require a systematic programme of research. Like the theory of mind account, it is to be hoped that, whether right or wrong, the central coherence theory will form a useful framework for thinking about autism in the future.

References American Psychological Association (1987). Diagnostic and Statistical Manual of Mental Disorders, 3rd revised edition (DSM-III-R). Washington, DC: American Psychological Association.

Autism: beyond "theory of mind"

77

Asperger, H. (1944). Die "autistischen Psychopathen" im Kindesalter. Archiv fur Psychiatrie und Nervenkrankheiten, 117, 76-136. Astington, J.W., Harris, P.L., & Olson, D.R. (Eds.) (1989). Developing theories of mind. New York: Cambridge University Press. Attwood, A.H., Frith, U., & Hermelin, B. (1988). The understanding and use of interpersonal gestures by autistic and Down's syndrome children. Journal of Autism and Developmental Disorders, 18, 241-257. Aurnhammer-Frith, U. (1969). Emphasis and meaning in recall in normal and autistic children. Language and Speech, 12, 29-38. Baron-Cohen, S. (1989a). The autistic child's theory of mind: A case of specific developmental delay. Journal of Child Psychology and Psychiatry, 30, 285-297. Baron-Cohen, S. (1989b). Perceptual role taking and protodeclarative pointing in autism. British Journal of Developmental Psychology, 7, 113-127. Baron-Cohen, S. (1992). Out of sight or out of mind? Another look at deception in autism. Journal of Child Psychology and Psychiatry, 33, 1141-1155. Baron-Cohen, S., Leslie, A.M., & Frith, U. (1985). Does the autistic child have a "theory of mind"? Cognition, 21, 37-46. Baron-Cohen, S., Leslie, A.M., & Frith, U. (1986). Mechanical, behavioural and intentional understanding of picture stories in autistic children. British Journal of Developmental Psychology, 4, 113-125. Baron-Cohen, S., Spitz, A., & Cross, P. (1993). Can children with autism recognise surprise? Cognition and Emotion, 7, 507-516. Baron-Cohen, S., Tager-Flusberg, H., & Cohen, D.J. (Eds.) (1993). Understanding other minds: Perspectives from autism. Oxford: Oxford University Press. Bartlett, F.C. (1932). Remembering: A study in experimental and social psychology. Cambridge, UK: Cambridge University Press. Bartlett, J.C., & Searcy, J. (1993). Inversion and configuration of faces. Cognitive Psychology, 25, 281-316. Bishop, D.V.M. (1993). Annotation. Autism, executive functions and theory of mind: A neuropsychological perspective. Journal of Child Psychology and Psychiatry, 34, 279-293. Bowler, D.M. (1992). "Theory of mind" in Asperger's syndrome. Journal of Child Psychology and Psychiatry, 33, 877-893. Byrne, R., & Whiten, A. (Eds.) (1988). Machiavellian intelligence: Social expertise and the evolution of intellect in monkeys, apes, and humans. Oxford: Clarendon Press. Cheney, D.L., & Seyfarth, R.M. (1990). How monkeys see the world. Chicago: University of Chicago Press. Cicchetti, D., & Cohen, D.J. (Eds.) (in press). Manual of developmental psychopathology (Vol. 1). New York: Wiley. Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason selection task. Cognition, 31, 187-276. Fodor, J.A. (1983). Modularity of mind. Cambridge, MA: MIT Press. Frith, U. (1970a). Studies in pattern detection in normal and autistic children: I. Immediate recall of auditory sequences. Journal of Abnormal Psychology, 76, 413-420. Frith, U. (1970b). Studies in pattern detection in normal and autistic children: II. Reproduction and production of color sequences. Journal of Experimental Child Psychology, 10, 120-135. Frith, U. (1989). Autism: Explaining the enigma. Oxford: Basil Blackwell. Frith, U. (1991). Translation and annotation of "Autistic psychopathy" in childhood, by H. Asperger. In U. Frith (Ed.), Autism and Asperger syndrome. Cambridge, UK: Cambridge University Press. Frith, U., Happe, F., & Siddons, F. (in press). Theory of mind and social adaptation in autistic, retarded and young normal children. Social Development. Frith, U., & Hermelin, B. (1969). The role of visual and motor cues for normal, subnormal and autistic children. Journal of Child Psychology and Psychiatry, 10, 153-163. Frith, U., Morton, J., & Leslie, A.M. (1991). The cognitive basis of a biological disorder: Autism. Trends in Neuroscience, 14, 433-438.

28

U. Frith, F. Happe"

Frith, U., & Snowling, M. (1983). Reading for meaning and reading for sound in autistic and dyslexic children. Journal of Developmental Psychology, 1, 329-342. Gillberg, C, & Coleman, M. (1992). The biology of the autistic syndromes. London: Mac Keith Press. Gottschaldt, K. (1926). Ueber den Einfluss der Erfahrung auf die Welt der Wahrnehmung von Figuren. Psychologische Forschung, 8, 261-317. Happe, F.G.E. (1991). Theory of mind and communication in autism. Unpublished Ph.D. thesis, University of London. Happe, F.G.E. (1993). Communicative competence and theory of mind in autism: A test of relevance theory. Cognition, 48, 101-119. Happe, F.G.E. (1994a). Annotation: Psychological theories of autism. Journal of Child Psychology and Psychiatry, 35, 215-229. Happe, F.G.E. (1994b). An advanced test of theory of mind: Understanding of story characters' thoughts and feelings by able autistic, mentally handicapped and normal children and adults. Journal of Autism and Developmental Disorders, 24, 1-24. Happe, F.G.E. (submitted). Theory of mind and IQ profiles in autism: A research note. Happe, F.G.E. (in preparation). Central coherence, block design errors, and global-local similarity judgement in autistic subjects. Happe, F., & Frith, U. (in press). Theory of mind in autism. In E. Schopler & G.B. Mesibov (Eds.), Learning and cognition in autism. New York: Plenum Press. Hart, C. (1989). Without reason: A family copes with two generations of autism. New York: Penguin Books. Hermelin, B., & O'Connor, N. (1967). Remembering of words by psychotic and subnormal children. British Journal of Psychology, 58, 213-218. Hermelin, B., & O'Connor, N. (1970). Psychological experiments with autistic children. Oxford: Pergamon. Hermelin, B., & O'Connor, N. (1986). Idiot savant calendrical calculators: Rules and regularities. Psychological Medicine, 16, 885-893. Hobson, R.P., Ouston, J., & Lee, T. (1988). What's in a face? The case of autism. British Journal of Psychology, 79, 441-453. Hughes, C.H., & Russell, J. (1993). Autistic children's difficulty with mental disengagement from an object: Its implications for theories of autism. Developmental Psychology, 29, 498-510. Kanner, L. (1943). Autistic disturbances of affective contact. Nervous Child, 2, 217-250. Kanner, L., & Eisenberg, L. (1956). Early infantile autism 1943-1955. American Journal of Orthopsychiatry, 26, 55-65. Kimchi, R., & Palmer, S.E. (1982). Form and texture in hierarchically constructed patterns. Journal of Experimental Psychology: Human Perception and Performance, 8, 521-535. Koffka, K. (1935). Principles of Gestalt psychology. New York: Harcourt Brace. Kohs, S.C. (1923). Intelligence measurement. New York: McMillan. Kramer, J.H., Kaplan, E., Blusewicz, M.J., & Preston, K.A. (1991). Visual hierarchical analysis of block design configural errors. Journal of Clinical and Experimental Neuropsychology, 13, 455-465. Landa, R., Folstein, S.E., & Isaacs, C. (1991). Spontaneous narrative-discourse performance of parents of autistic individuals. Journal of Speech and Hearing Research, 34, 1339-1345. Langdell, T. (1978). Recognition of faces: An approach to the study of autism. Journal of Child Psychology and Psychiatry, 19, 255-268. Leekam, S., & Perner, J. (1991). Does the autistic child have a metarepresentational deficit? Cognition, 40, 203-218. Leslie, A.M. (1987). Pretence and representation: The origins of "Theory of Mind". Psychological Review, 94, 412-426. Leslie, A.M. (1988). Some implications of pretence for mechanisms underlying the child's theory of mind. In J.W. Astington, P.L. Harris, & D.R. Olson (Eds.), Developing theories of mind. New York: Cambridge University Press. Leslie, A.M., & Thaiss, L. (1992). Domain specificity in conceptual development: Evidence from autism. Cognition, 43, 225-251.

Autism: beyond "theory of mind"

29

Lockyer, L., & Rutter, M. (1970). A five to fifteen year follow-up study of infantile psychosis: IV. Patterns of cognitive ability. British Journal of Social and Clinical Psychology, 9, 152-163. McDonnell, J.T. (1993). News from the Border: A mother's memoir of her autistic son. New York: Ticknor & Fields. Morton, J., & Frith, U. (in press). Causal modelling: A structural approach to developmental psychopathology. In D. Cicchetti & D.J. Cohen (Eds.), Manual of Developmental Psychopathology (Vol. 1, Ch. 13). New York: Wiley. Mottron, L., & Belleville, S. (1993). A study of perceptual analysis in a high-level autistic subject with exceptional graphic abilities. Brain and Cognition, 23, 279-309. Navon, D. (1977). Forest before trees: The precedence of global features in visual perception. Cognitive Psychology, 9, 353-383. Newson, E., Dawson, M., & Everard, P. (1984). The natural history of able autistic people: Their management and functioning in social context. Summary of the report to DHSS in four parts. Communication, 18, 1-4; 19, 1-2. Norris, D. (1990). How to build a connectionist idiot (savant). Cognition, 35, 277-291. O'Connor, N., & Hermelin, B. (1984). Idiot savant calendrical calculators: Maths or memory. Psychological Medicine, 14, 801-806. Ozonoff, S., Pennington, B.F., & Rogers, S.J. (1991). Executive function deficits in high-functioning autistic children: Relationship to theory of mind. Journal of Child Psychology and Psychiatry, 32, 1081-1106. Ozonoff, S., Rogers, S.J., & Pennington, B.F. (1991). Asperger's syndrome: Evidence of an empirical distinction from high-functioning autism. Journal of Child Psychology and Psychiatry, 32, 1107-1122. Park, C.C. (1967). The siege: The battle for communication with an autistic child. Harmondsworth, UK: Penguin Books. Perner, J. (1991). Understanding the representational mind. Cambridge, MA: MIT Press. Perner, J., Frith, U., Leslie, A.M., & Leekam, S.R. (1989). Exploration of the autistic child's theory of mind: Knowledge, belief, and communication. Child Development, 60, 689-700. Perner, J., & Wimmer, H. (1985). "John thinks that Mary thinks that . . .": Attribution of second-order beliefs by 5-10 year old children. Journal of Experimental Child Psychology, 39, 437-471. Phillips, W. (1993). Understanding intention and desire by children with autism. Unpublished Ph.D. thesis, University of London. Premack, D., & Woodruff, G. (1978). Does the chimpanzee have a theory of mind? Behavioural and Brain Sciences, 4, 515-526. Prior, M.R. (1979). Cognitive abilities and disabilities in infantile autism: A review. Journal of Abnormal Child Psychology, 7, 357-380. Rhodes, G., Brake, S., & Atkinson, A.P. (1993). What's lost in inverted faces? Cognition, 47, 25-57. Rimland, B., & Hill, A.L. (1984). Idiot savants. In J. Wortis (Ed.), Mental retardation and developmental disabilities (vol. 13, pp. 155-169). New York: Plenum Press. Russell, J. (1992). The theory-theory: So good they named it twice? Cognitive Development, 7, 485-519. Rutter, M. (1987). The role of cognition in child development and disorder. British Journal of Medical Psychology, 60, 1-16. Schopler, E., & Mesibov, G.B. (Eds.) (1987). Neurobiological issues in autism. New York: PlenumPress. Shah, A., & Frith, U. (1983). An islet of ability in autistic children: A research note. Journal of Child Psychology and Psychiatry, 24, 613-620. Shah, A., & Frith, U. (1993). Why do autistic individuals show superior performance on the Block Design task? Journal of Child Psychology and Psychiatry, 34, 1351-1364. Snowling, M., & Frith, U. (1986). Comprehension in "hyperlexic" readers. Journal of Experimental Child Psychology, 42, 392-415. Sodian, B., & Frith, U. (1992). Deception and sabotage in autistic, retarded and normal children. Journal of Child Psychology and Psychiatry, 33, 591-605. Sperber, D., & Wilson, D. (1986). Relevance: Communication and cognition. Oxford: Blackwell.

30

U. Frith, F. Happ

o

X OTC

o

Subjective Reward Magniti

CD X5

I

i i i_ , i 0.1 0.2 0.5 1 2 Trai n Duration (s)

Fig. 2. The measurements of the subjective reward magnitude at different train durations made by the equipreference method are compared to the measurements made by the direct method (with the rates of reward the same on both levers). The curve was computed by a smoothing routine from the complete data set shown in Fig. 3. The approximate agreement between the two sets of measurements validates the assumption that the subjective rate of reward is proportional to the objective rate. This is a key assumption in the equipreference method, but this assumption is unnecessary in the direct method, because, in the direct method, the relative rate of reward is held constant. In the data shown here, the relative rate of reward was 1:1. Data from Mark and Gallistel (1993).

Foraging for brain stimulation

59

reward combines multiplicatively with subjective magnitude to determine return, then the relative rate of reward should act as a simple scaling factor when subjective reward magnitude is measured by the direct method. Suppose that when the rates of reward on the two levers are equal, the rat has a 4:1 preference for the reward generated by a 1 s train over the reward generated by a 0.5 s train. The direct method takes this to mean that the subjective reward from the 1 s train is four times bigger than the subjective reward from the 0.5 s train. Suppose we repeat the measurements with the schedules of reward adjusted so that the 0.5 s reward comes twice as often as the 1 s reward. The rat's preference (time allocation ratio) for the 1 s lever should be reduced by a factor of two. It should allocate only twice as much time to the 1 s lever. Suppose we repeat the measurements with the schedules adjusted so that the 0.5 s reward comes only half as often as the 1 s reward. The rat should now allocate eight times as much time to the 1 s lever (a factor of two increase in its preference). And so on. At a given setting of the relative rates of reward, Mark and Gallistel (1993) determined the rat's time allocation ratio as a function of the train duration on one lever, keeping the reward on the other lever constant. The time allocation ratios are the direct measures of the subjective magnitude of the variable reward. A set of these time allocation ratios, one for each duration of the variable reward, gives the function relating subjective reward magnitude to the duration of the train of stimulation. If the relative rate of reward acts as a scaling factor, then the functions we obtained at different relative rates of reward should be superimposable, provided we correct for the difference in the scale of measurement. To correct for differences in the scale of measurement, we multiplied the time allocation ratios from a given session by the inverse of the rate ratio in effect for that session. Fig. 3 shows that, after rescaling, the measurements made at different relative rates of reward superimpose. This validates the assumption that subjective rate of reward combines multiplicatively with subjective reward magnitude to determine the subjective return from a lever. The greatest value of these measurement experiments lies in what they reveal about the decision process, the computational process that uses these psychological variables (subjective magnitude and subjective rate) to determine how the animal will behave. In the course of validating the measurements, we have established that the decision process multiplies subjective reward magnitude by subjective rate of reward to determine the subjective value of the lever. Thus, we have isolated a simple computational process, where we can control the values of the variables that enter into the computation by direct electrical stimulation of a pathway in the central nervous system. Moreover, we have determined how those values depend on the strength and the duration of the barrage of action potentials produced by the stimulation. The subjective magnitude of the reward is a steep sigmoidal function of the strength of the neural signal (Gallistel & Leon, 1991;

60

C. Gallistel 2

nitu (cost to offerer) (Cosmides & Tooby, 1989). To become grammatical, the context must cause the violated constraint to be satisfied. For example, recategorizing the gum wrapper as something extremely valuable (potentially justifying the $1000 payment) would do this: the statement seems sensible if you are told that the speaker is a spy who knows the gum wrapper has a microdot with the key for breaking an enemy code. n The term "innate" means different things to different scientific communities, but no person who uses the term means "immune to every environmental perturbation". UG is innate in the following sense: its intricate internal organization is the product of our species' genetic endowment in the same way that the internal organization of the eye is. Its neurological development is buffered against most naturally occurring variations in the physical and social environment. Certain environmental conditions are necessary to trigger the development of UG, but these conditions are not the source of its internal organization. As a result, all normal human beings raised in reasonably normal environments develop the same UG (e.g., Pinker, 1994). For an extensive discussion of how natural selection structures the relationships among genotype, phenotype and environment in development, see Tooby and Cosmides (1992).

98

L. Cosmides, J. Tooby

Instinct blindness UG is a small corner of hypothesis space; there are an indefinitely large number of grammars that are not variants of UG. To explain the fact that all natural languages fall within the bounds of UG, one must first realize that UG exists. To realize that it exists, one must realize that there are alternative grammars. But this last step is where our imagination stumbles. The language instinct structures our thought so powerfully that alternative grammars are difficult to imagine. This is not an incidental feature of the language instinct; it is the language acquisition device's (LAD) principal adaptive function.12 Any set of utterances a child hears is consistent with an infinite number of possible grammars, but only one of them is the grammar of its native language. A content-free learning mechanism would be forever lost in hypothesis space. The LAD is an adaptation to combinatorial explosion: by restricting the child's grammatical imagination to a very small subset of hypothesis space - hypotheses consistent with the principles of UG - it makes language acquisition possible. Its function is to generate grammatical inferences consistent with UG without simultaneously generating inconsistent ones. To do this, the LAD's structure must make alternative grammars literally unimaginable (at least by the language faculty). This is good for the child learning language, but bad for the cognitive scientist, who needs to imagine these unimaginable grammars. Forming the plural through mirror reversal - so that the plural of "cat" is "tac" - is a rule in an alternative grammar. No child considers this possibility; the LAD cannot generate this rule. The cognitive scientist needs to know this, however, in order to characterize UG and produce a correct theory of the LAD's cognitive structure. UG is what, an algorithm is how. A proposed algorithm can be ruled out, for example, if formal analyses reveal that it produces both the mirror reverse rule and the "add 's' to a stem" rule. Alternative grammars - and hence Universal Grammar - were difficult to discover because circuits designed to generate only a small subset of all grammatical inferences in the child also do so in the linguist. This property of the language instinct is crucial to its adaptive function. But it caused a form of theoretical blindness in linguists, which obstructed the discovery of UG and of the language instinct itself. One can think of this phenomenon as instinct blindness. Discovering a grammar of social reasoning is likely to prove just as difficult as discovering the grammar of a language, and for exactly the same reasons. Yet 12 As a side-effect, it can also solve problems that played no causal role in its selective history. For example, the LAD was not designed to support writing, but its properties made the design and spread of this cultural invention possible.

Beyond intuition and instinct blindness

99

there is no field, parallel to linguistics, that is devoted to this task; indeed, very few individuals even recognize the need for such a grammar, let alone such a field (for exceptions, see Cosmides, 1985; Cosmides & Tooby, 1989, 1992; Fiske, 1991; Jackendoff, 1992). Our intuitions blind us not only to the existence of instincts, but to their complexity. The phenomenal experience of an activity as "easy" or "natural" often leads scientists to assume that the processes that give rise to it are simple. Legend has it that in the early days of artificial intelligence, Marvin Minsky assigned the development of machine vision to a graduate student as a summer project. This illusion of simplicity hampered vision research for years: . . . in the 1960s almost no one realized that machine vision was difficult. The field had to go through [a series of fiascoes] before it was at last realized that here were some problems that had to be taken seriously. The reason for this misperception is that we humans are ourselves so good at vision. (Marr, 1982, p. 16)

Phenomenally, seeing seems simple. It is effortless, automatic, reliable, fast, unconscious and requires no explicit instruction. But seeing is effortless, automatic, reliable, fast, and unconscious precisely because there is a vast array of complex, dedicated computational machinery that makes this possible. Most cognitive scientists don't realize it, but they are grossly underestimating the complexity of our central processes. To find someone beautiful, to fall in love, to feel jealous, to experience moral outrage, to fear disease, to reciprocate a favor, to initiate an attack, to deduce a tool's function from its shape - and a myriad other cognitive accomplishments - can seem as simple and automatic and effortless as opening your eyes and seeing. But this apparent simplicity is possible only because there is a vast array of complex computational machinery supporting and regulating these activities. The human cognitive architecture probably embodies a large number of domain-specific "grammars", targeting not just the domain of social life, but also disease, botany, tool-making, animal behavior, foraging and many other situations that our hunter-gatherer ancestors had to cope with on a regular basis. Research on the computational machinery responsible for these kinds of inferences, choices and preferences - especially the social ones-is almost totally absent in the cognitive sciences. This is a remarkable omission, from an evolutionary point of view. Instinct blindness is one culprit; extreme and unfounded claims about cultural relativity is another (e.g., Brown, 1991; Sperber, 1982; Tooby & Cosmides, 1992).

Anthropological malpractice As a result of the rhetoric of anthropologists, most cognitive researchers have, as part of their standard intellectual furniture, a confidence that cultural relativity

L. Cosmides, J. Tooby

100

is an empirically established finding of wide applicability (see discussion of the Standard Social Science Model in Tooby & Cosmides, 1992). Consequently, most scientists harbor the incorrect impression that there is no "Universal Grammar" of social reasoning to be discovered. According to this view, a grammar of social reasoning might exist in each culture, but these grammars will differ dramatically and capriciously from one culture to the next. In its most extreme form, the relativist position holds that the grammars of different cultures are utterly incommensurate - that there is no transformation that can map the rules of one onto the rules of another. If so, then these rules cannot be expressions of an underlying UG of social reasoning. Among anthropologists, however, cultural relativism is an interpretation imposed as an article of faith - not a conclusion based on scientific data (Brown, 1991; Sperber, 1982; Tooby & Cosmides, 1992).13 Indeed, Maurice Bloch, a prominent member of the field, has complained that it is the "professional malpractice of anthropologists to exaggerate the exotic character of other cultures" (Bloch, 1977). To some degree, this is a self-legitimizing institutional pressure: why go long distances to study things that could be studied at home (Brown, 1991)? More importantly, however, anthropologists are just as oblivious to what is universally natural for the human mind as the rest of us. Their attention is drawn to what differs from culture to culture, not what is absent from all cultures or what differs from species to species. Drawing on their cognitive instincts, they understand, automatically and without reflection, much of what happens in other cultures. They know they can work out exchanges without language, or see a smile, a shared look, or an aggressive gesture and infer its meaning and its referent. Indeed, they operate within a huge set of implicit panhuman assumptions that allow them to decode the residue of human life that does differ from place to place (Sperber, 1982; Tooby & Cosmides, 1992). The notion of universal human reasoning instincts - including social reasoning instincts - is completely compatible with the ethnographic record. It is more than empirically reasonable; it is a logical necessity, for the reasons discussed above. Indeed, without universal reasoning instincts, the acquisition of one's "culture" would be literally impossible, because one wouldn't be able to infer which representations, out of the infinite universe of possibilities, existed in the minds of other members of the culture (Boyer, 1994; Chomsky, 1980; Sperber, 1985, 1990; Tooby & Cosmides, 1992). Instinct blindness is a side-effect of any instinct whose function is to generate some inferences or behaviors without simultaneously generating others. This is a

13

For a history and discussion of how unsupported relativist claims gained widespread acceptance in the social sciences, see Brown (1991) and Tooby and Cosmides (1992).

Beyond intuition and instinct blindness

101

very general property of instincts, because combinatorial explosion is a very general selection pressure (for discussion, see Tooby & Cosmides, 1992). The fact that human instincts are difficult for human minds to discover is a side-effect of their adaptive function. Many aspects of the human mind can't be seen by the naked "I" - by intuition unaided by theory. A good theory rips away the veil of naturalness and familiarity that our own minds create, exposing computational problems whose existence we never even imagined. The cognitive sciences need theoretical guidance that is grounded in something beyond intuition. Otherwise, we're flying blind.

Corrective lenses There are various ways of overcoming instinct blindness. One of the most common is the study of non-human minds that differ profoundly from our own - animal minds and electronic minds, broody hens and AI programs. Linguists were awakened to the existence of alternative grammars by the creation of computer "languages", which are not variants of UG. These languages "made the natural seem strange", inspiring linguists to generate even stranger grammars. To do this, they had to escape the confines of their intuitions, which they did through the use of mathematical logic and the theory of computation. In William James's terms, they debauched their minds with learning. The study of animal behavior is another time-honored method for debauching the mind-the one used by William James himself. Hermaphroditic worms, colonies of ant sisters who come in three "genders" (sterile workers, soldiers, queens), male langur monkeys who commit systematic infanticide when they join a troop, flies who are attracted to the smell of dung, polyandrous jacanas who mate with a male after breaking the eggs he was incubating for a rival female, fish who change sex when the composition of their social group changes, female praying mantises who eat their mate's head while copulating with him -other animals engage in behaviors that truly are exotic by human standards. Human cultural variation is trivial in comparison. Observing behaviors caused by alternative instincts jars us into recognizing the specificity and multiplicity of our own instincts. Observations like these tell us what we are not, but not what we are. That's why theoretical biology is so important. It provides positive theories of what kinds of cognitive programs we should expect to find in species that evolved under various ecological conditions: theories of what and why. Evolutionary biology's formal theories are powerful lenses that correct for instinct blindness. In their focus, the intricate outlines of the mind's design stand out in sharp relief.

102

L. Cosmides, J. Tooby

References Atran, S. (1990). The cognitive foundations of natural history. New York: Cambridge University Press. Axelrod, R. (1984). The evolution of cooperation. New York: Basic Books. Axelrod, R., & Hamilton, W.D. (1981). The evolution of cooperation. Science, 211, 1390-1396. Barkow, J., Cosmides, L., & Tooby, J. (Eds.) (1992). The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Baron-Cohen, S., Leslie, A., & Frith, U. (1985). Does the autistic child have a "theory of mind"? Cognition, 21, 37-46. Bloch, M. (1977). The past and the present in the present. Man, 12, 278-292. Boyd, R. (1988). Is the repeated prisoner's dilemma a good model of reciprocal altruism? Ethology and Sociobiology, 9, 211-222. Boyer, P. (1994). The naturalness of religious ideas. Berkeley: University of California Press. Brothers, L. (1990). The social brain: A project for integrating primate behavior and neurophysiology in a new domain. Concepts in Neuroscience, 1, 27-51. Brown, A. (1990). Domain-specific principles affect learning and transfer in children. Cognitive Science, 14, 107-133. Brown, D.E. (1991). Human universals. New York: McGraw-Hill. Carey, S. (1985). Constraints on semantic development. In J. Mehler & R. Fox (Eds.), Neonate cognition (pp. 381-398). Hillsdale, NJ: Erlbaum. Carey, S., & Gelman, R. (Eds.) (1991). The epigenesis of mind. Hillsdale, NJ: Erlbaum. Cheney, D.L., & Seyfarth, R. (1990). How monkeys see the world. Chicago: University of Chicago Press. Cheng, P., Holyoak, K., Nisbett, R., & Oliver, L. (1986). Pragmatic versus syntactic approaches to training deductive reasoning. Cognitive Psychology, 18, 293-328. Chomsky, N. (1980). Rules and representations. New York: Columbia University Press. Clutton-Brock, T.H., & Harvey, P. (1979). Comparison and adaptation. Proceedings of the Royal Society, London B, 205, 547-565. Cook, E.W., III, Hodes, R.L., & Lang, P.J. (1986). Preparedness and phobia: Effects of stimulus content on human visceral conditioning. Journal of Abnormal Psychology, 95, 195-207. Cook, M., & Mineka, S. (1989). Observational conditioning of fear to fear-relevant versus fearirrelevant stimuli in rhesus monkeys. Journal of Abnormal Psychology, 98, 448-459. Cosmides, L. (1985). Deduction or Darwinian algorithms? An explanation of the "elusive" content effect on the Wason selection task. Doctoral dissertation, Department of Psychology, Harvard University. University Microfilms #86-02206. Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason selection task. Cognition, 31, 187-276. Cosmides, L., & Tooby, J. (1987). From evolution to behavior: Evolutionary psychology as the missing link. In J. Dupre (Ed.), The latest on the best: Essays on evolution and optimality. Cambridge, MA: MIT Press. Cosmides, L., & Tooby, J. (1989). Evolutionary psychology and the generation of culture, Part II. Case study: A computational theory of social exchange. Ethology and Sociobiology, 10, 51-97. Cosmides, L., & Tooby, J. (1992). Cognitive adaptations for social exchange. In J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Cosmides, L., & Tooby, J. (1994). Origins of domain specificity: The evolution of functional organization. In S. Gelman & L. Hirschfeld (Eds.), Mapping the mind: Domain specificity in cognition and culture. New York: Cambridge University Press. Cosmides, L., & Tooby, J. (in press). Are humans good intuitive statisticians after all? Rethinking some conclusion of the literature on judgment under uncertainty. Cognition. Daly, M., & Wilson, M. (1983). Sex, evolution and behavior. Boston: Wadsworth. Daly, M., & Wilson, M. (1988). Homicide. Hawthorne, NY: Aldine de Gruyter. Daly, M., & Wilson, M. (1994). Discriminative parental solicitude and the relevance of evolutionary

Beyond intuition and instinct blindness

103

models to the analysis of motivational systems. In M. Gazzaniga (Ed.), The cognitive neurosciences. Cambridge, MA: MIT Press. Dawkins, R. (1982). The extended phenotype. Oxford: W.H. Freeman. Dawkins, R. (1986). The blind watchmaker. New York: Norton. de Waal, F.B.M., & Luttrell, L.M. (1988). Mechanisms of social reciprocity in three primate species: Symmetrical relationship characteristics or cognition? Ethology and Sociobiology, 9, 101-118. Fischer, E.A. (1988). Simultaneous hermaphroditism, tit-for-tat, and the evolutionary stability of social systems. Ethology and Sociobiology, 9, 119-136. Fiske, A.P. (1991). Structures of social life: The four elementary forms of human relations. New York: Free Press. Fodor, J.A. (1983). The modularity of mind. Cambridge, MA: MIT Press. Frith, U. (1989). Autism: Explaining the enigma. Oxford: Blackwell. Gallistel, C.R. (1990). The organization of learning. Cambridge, MA: MIT Press. Gallistel, C.R., Brown, A.L., Carey, S., Gelman, R., & Keil, F.C. (1991). Lessons from animal learning for the study of cognitive development. In S. Carey & R. Gelman (Eds.), The epigenesis of mind. Hillsdale, NJ: Erlbaum. Garcia, J. (1990). Learning without memory. Journal of Cognitive Neuroscience, 2, 287-305. Gelman, S., & Hirschfeld, L. (Eds.) (1994). Mapping the mind: Domain specificity in cognition and culture. New York: Cambridge University Press. Gigerenzer, G., Hoffrage, U., & Kleinbolting, H. (1991). Probabilistic mental models: A Brunswikean theory of confidence. Psychological Review, 98, 506-528. Gigerenzer, G., & Hug, K. (1992). Domain-specific reasoning: Social contracts, cheating and perspective change. Cognition, 43, 127-171. Gould, J.L. (1982). Ethology: The mechanisms and evolution of behavior. New York: Norton. Gould, S.J., & Lewontin, R.C. (1979). The spandrels of San Marco and the Panglossian paradigm: A critique of the adaptationist programme. Proceedings of the Royal Society, London B, 205, 581-598. Jackendoff, R. (1992). Languages of the mind. Cambridge, MA: MIT Press. James, W. (1890). Principles of psychology. New York: Henry Holt. Keil, F.C. (1989). Concepts, kinds, and cognitive development. Cambridge, MA: MIT Press. Krebs, J.R., & Davies, N.B. (1987). An introduction to behavioural ecology. Oxford: Blackwell Scientific Publications. Leslie, A.M. (1987). Pretense and representation: The origins of "theory of mind". Psychological Review, 94, 412-426. Leslie, A.M. (1988). The necessity of illusion: Perception and thought in infancy. In L. Weiskrantz (Ed.), Thought without language (pp. 185-210). Oxford: Clarendon Press. Manktelow, K.I., & Evans, J. St.B.T. (1979). Facilitation of reasoning by realism: Effect or non-effect? British Journal of Psychology, 70, 477-488. Manktelow, K.I., & Over, D.E. (1990). Deontic thought and the selection task. In K.J. Gilhooly, M.T.G. Keane, R.H. Logie, & G. Erdos (Eds.), Lines of thinking (Vol. 1). Chichester: Wiley. Markman, E.M. (1989). Categorization and naming in children: Problems of induction. Cambridge, MA: MIT Press. Markman, E. (1990). Constraints children place on word meanings. Cognitive Science, 14, 57-77. Marler, P. (1991). The instinct to learn. In S. Carey & R. Gelman (Eds.), The epigenesis of mind. Hillsdale, NJ: Erlbaum. Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. San Francisco: Freeman. Mayr, E. (1983). How to carry out the adaptationist program? The American Naturalist, 121, 324-334. McDougall, W. (1908/1916). Introduction to social psychology. Boston: John W. Luce. Mineka, S., & Cook, M. (1988). Social learning and the acquisition of snake fear in monkeys. In T.R. Zentall & B.G. Galef (Eds.), Social learning: Psychological and biological perspectives, (pp. 51-73). Hillsdale, NJ: Erlbaum. Ohman, A., Dimberg, U., & Ost, L.G. (1985). Biological constraints on the fear response. In S. Reiss & R. Bootsin (Eds.), Theoretical issues in behavior therapy (pp. 123-175). New York: Academic Press.

104

L. Cosmides, J. Tooby

Ohman, A., Eriksson, A., & Olofsson, A. (1975). One-trial learning and superior resistance to extinction of autonomic responses conditioned to potentially phobic stimuli. Journal of Comparative and Physiological Psychology, 88, 619-627. Pinker, S. (1994). The language instinct. New York: Morrow. Pinker, S., & Bloom, P. (1990). Natural language and natural selection. Behavioral and Brain Sciences, 13, 707-727. Profet, M. (1992). Pregnancy sickness as adaptation: A deterrent to maternal ingestion of teratogens. In J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Pylyshyn, Z.W. (Ed.) (1987). The robot's dilemma: The frame problem in artificial intelligence. Norwood, NJ: Ablex. Real, L.A. (1991). Animal choice behavior and the evolution of cognitive architecture. Science, 253, 980-986. Rozin, P. (1976). The evolution of intelligence and access to the cognitive unconscious. In J.M. Sprague & A.N. Epstein (Eds.), Progress in psychobiology and physiological psychology. New York: Academic Press. Shepard, R.N. (1981). Psychophysical complementarity. In M. Kubovy & J. Pomerantz (Eds.), Perceptual organization. Hillsdale, NJ: Erlbaum. Shepard, R.N. (1987). Evolution of a mesh between principles of the mind and regularities of the world. In J. Dupre (Ed.), The latest on the best: Essays on evolution and optimality. Cambridge, MA: MIT Press. Sherry, D.F., & Schacter, D.L. (1987). The evolution of multiple memory systems. Psychological Review, 94, 439-454. Smuts, B. (1986). Sex and friendship in baboons. Hawthorne: Aldine. Spelke, E.S. (1988). The origins of physical knowledge. In L. Weiskrantz (Ed.), Thought without language (pp. 168-184). Oxford: Clarendon Press. Spelke, E. (1990). Principles of object perception. Cognitive Science, 14, 29-56. Sperber, D. (1975) Rethinking symbolism, transl. Alice Morton. Cambridge, UK: Cambridge University Press. Sperber, D. (1982). On anthropological knowledge. Cambridge, UK: Cambridge University Press. Sperber, D. (1985). Anthropology and psychology: Towards an epidemiology of representations. Man (N.S.), 20, 73-89. Sperber, D. (1990). The epidemiology of beliefs. In C. Fraser & G. Geskell (Eds.), Psychological studies of widespread beliefs. Oxford: Clarendon Press. Sperber, D. (1994). The modularity of thought and the epidemiology of representations. In S. Gelman & L. Hirschfeld, (Eds.), Mapping the mind: Domain specificity in cognition and culture. New York: Cambridge University Press. Symons, D. (1979). The evolution of human sexuality. New York: Oxford University Press. Symons, D. (1992). On the use and misuse of Darwinism in the study of human behavior. In J. Barkow, L. Cosmides, & J. Tooby, (Eds.), The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Tooby, J., & Cosmides, L. (1989). The logic of threat: Evidence for another cognitive adaptation? Paper presented at the Human Behavior and Evolution Society, Evanston, IL. Tooby, J., & Cosmides, L. (1990a). The past explains the present: Emotional adaptations and the structure of ancestral environments. Ethology and Sociobiology, 11, 375-424. Tooby, J., & Cosmides, L. (1990b). On the universality of human nature and the uniqueness of the individual: The role of genetics and adaptation. Journal of Personality, 58, 17-67. Tooby, J., & Cosmides, L. (1992). The psychological foundations of culture. In J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Tooby J., & DeVore, I. (1987). The reconstruction of hominid behavioral evolution through strategic modeling. In W. Kinzey (Ed.), Primate models of hominid behavior. New York: SUNY Press. Trivers, R.L. (1971). The evolution of reciprocal altruism. Quarterly Review of Biology, 46, 35-57. Wason, P. (1966). Reasoning. In B.M. Foss (Ed.), New horizons in psychology. Harmondsworth: Penguin.

Beyond intuition and instinct blindness

105

Wason, P. (1983). Realism and rationality in the selection task. In J. St.B.T. Evans (Ed.), Thinking and reasoning: Psychological approaches. London: Routledge & Kegan Paul. Wason, P., & Johnson-Laird, P.N. (1972). Psychology of reasoning: Structure and content. London: Batsford. Wilkinson, G.S. (1988). Reciprocal altruism in bats and other mammals. Ethology and Sociobiology, 9, 85—100. Wilkinson, G.S. (1990). Food sharing in vampire bats. Scientific American, February, 76-82. Williams, G.C. (1966). Adaptation and natural selection: A critique of some current evolutionary thought. Princeton: Princeton University Press. Williams, G.C. (1985). A defense of reductionism in evolutionary biology. Oxford surveys in evolutionary biology, 2, 1-27. Williams, G.C, & Nesse, R.M. (1991). The dawn of Darwinian medicine. Quarterly Review of Biology, 66, 1-22. Wilson, M., & Daly, M. (1992). The man who mistook his wife for a chattel. In J. Barkow, L. Cosmides, & J. Tooby (Eds.), The adapted mind: Evolutionary psychology and the generation of culture. New York: Oxford University Press. Wynn, K. (1992). Addition and subtraction by human infants. Nature, 358, 749-750.

6 Why should we abandon the mental logic hypothesis? Luca Bonatti* Laboratoire de Sciences Cognitives et Psycholinguistique, 54, Boulevard Raspail, 75006 Paris, France Philosophy Department, Rutgers University, PO Box 270, New Brunswick, NJ 08903-0270, USA

Abstract Two hypotheses on deductive reasoning are under development: mental logic and mental models. It is often accepted that there are overwhelming arguments to reject the mental logic hypothesis. I revise these arguments and claim that they are either not conclusive, or point at problems which are troublesome for the mental model hypothesis as well.

1. Introduction An old and venerable idea holds that logic is concerned with discovering or illuminating the laws of thought. Its psychological corollary is that a system of logic in the mind underlines our thinking processes. This thesis fits very well with representational views of the mind according to which cognitive processes are largely proof-theoretical. Within such a framework, it is a thesis about the structure of the vehicle of internal representations. In a nutshell, it holds that reasoning consists of operations on mental representations, according to logical rules implemented in procedures activated by the forms of the mental representations. Even if the thesis loomed around for centuries, there is still little convincing psychological evidence of the existence of a mental logic. Such evidence has mostly been accumulated in the last few years, and almost exclusively concerns propositional reasoning (Braine, Reiser & Rumain, 1984; Lea, O'Brien, Fisch, Noveck & Braine, 1990; Rips, 1983). * Correspondence to: L. Bonatti, Laboratoire de Sciences Cognitives et Psycholinguistique, 54, Boulevard Raspail, 75006 Paris, France. The author is indebted to Martin Braine, Emmanuel Dupoux, Jerry Fodor, Jacques Mehler, and Christophe Pallier for comments on a first draft of this paper.

110

L. Bonatti

In the same years in which some results were beginning to appear, mental logic has been seriously challenged by an alternative - mental models - mostly due to the work of Johnson-Laird and his collaborators. Both hypotheses share the basic geography of cognition: also the mental models hypothesis is (inter alia) about the nature of the internal representations of deductive processes. They differ, however, on their supposed nature. Roughly, the mental model hypothesis claims that understanding a text consists of the manipulation of tokens representing concrete samples of entities in the world, and reasoning consists of the construction of alternative arrangements of tokens. No abstract rules should be needed to accomplish deduction. Thus, at least at first blush, while mental logic seems naturally to require a language of thought on whose formulas abstract rules apply, mental models seem to be able to dispense with it and substitute analog simulations for discrete manipulation of propositional-like objects (McGinn 1989). Originally, crucial aspects of the new hypothesis were left vague, and both its exact status and the feasibility of its claims were a puzzle (Boolos, 1984; Rips, 1986). What precisely a mental model is seemed to be a question of secondary importance, if compared to the big revolution introduced by the theory. Only recently has a substantial effort of formal clarification been undertaken (especially in Johnson-Laird & Byrne, 1991 and Johnson-Laird, Byrne, & Schaeken, 1992), but the task is still far from being accomplished (Bonatti, in press; Hodges, 1993). Nevertheless, the hypothesis had an enormous success, to the point that probably the words "mental models" are second only to "generative grammar" for their consequences within the cognitive science community. In a very short time, among psychologists an almost unanimous consensus has been reached on the death of mental logic and on the fact that reasoning is carried out by constructing mental models; nowadays the group of psychologists who doubt of the truth of the mental model theory is on the verge of extinction. A good part of this sweeping success, vagueness notwithstanding, is due to the impressive list of problems the new hypothesis promised to solve. Let me list them. Mental models would: (1) provide a general theory of deductive reasoning (Johnson-Laird, 1983a; Johnson-Laird & Bara, 1984, p. 3; Johnson-Laird & Byrne, 1991, p. x), and,' in particular (la) explain propositional reasoning (Johnson-Laird & Byrne, 1991, 1993; Johnson-Laird, Byrne, & Schaeken, 1992); (lb) explain relational reasoning (Johnson-Laird, 1983b; Johnson-Laird & Byrne 1989, 1991, 1993); (lc) explain the figural effect in reasoning (Johnson-Laird & Bara, 1984; Johnson-Laird & Byrne, 1991, Ch. 6); (Id) explain syllogistic reasoning (Johnson-Laird, 1983a, Ch. 5; Johnson-Laird & Bara, 1984; Johnson-Laird & Byrne, 1991), including individual differences

Why should we abandon the mental logic hypothesis?

Ill

(Johnson-Laird, 1983a, pp. 117-121) and the belief bias effect (JohnsonLaird & Byrne, 1991, pp. 125-126; Oakill, Johnson-Laird, & Garnham, 1989); (le) explain reasoning with single and multiple quantifiers (Johnson-Laird, 1983a; Johnson-Laird & Byrne, 1991; Johnson-Laird, Byrne, & Tabossi, 1989); (2) explain how logical reasoning is performed without logic (Byrne, 1991; Johnson-Laird, 1983a, Ch. 6, 1983b; Johnson-Laird & Byrne, 1991); (3) account for a vast series of linguistic phenomena, such as anaphors, definite and indefinite descriptions, pronouns and plausibility effects in language processing (Johnson-Laird, 1983a; Garnham, 1987); (4) offer a theory of the structure of discourse (Johnson-Laird, 1983a, pp. 370-371; Garnham, 1987); (5) explain the difference between implicit and explicit inferences (JohnsonLaird, 1983a, Ch. 6); (6) "solve the central paradox of how children learn to reason" (Johnson-Laird, 1983a, p. 45); (7) explain content effects in reasoning (Byrne, 1991, p. 77); (8) offer an explanation of meaning (Johnson-Laird, 1983a, p. 397; McGinn, 1989); (9) "readily cope with the semantics of propositional attitudes" (Johnson-Laird, 1983a, p. 430) and solve the problems presented by them (Johnson-Laird, 1983a, pp. 430-436); (10) provide a solution to the controversy on the problem of human rationality (Johnson-Laird & Byrne, 1993, p. 332); (11) solve the problem of how words relate to the world (Johnson-Laird, 1983a, p. 402, 1989, pp. 473-474, 489; Garnham, 1987; McGinn, 1989); (12) elucidate the nature of self-awareness and consciousness (Johnson-Laird, 1983a, pp. xi; Ch. 16). Even the most benevolent reader, when confronted with a theory so rich in both philosophical consequences and empirical power, should have at least felt inclined to raise her critical eyebrows. Nevertheless, critical voices were confined to a "small chorus of dissenters", almost all tied to the "ardent advocates of rule theories" (Johnson-Laird & Byrne, 1991, p. ix). In fact, with some patience and time, I think it can be shown that all the philosophical advantages claimed for mental models are unsupported propaganda, and that most of the psychological evidence is much less firm than generally admitted. But showing it is quite a long task. Another source of support for the mental model hypothesis came from a parallel series of arguments to the conclusion that the mental logic hypothesis is doomed to failure. In this paper, I will confine myself to a modest task. I will

112

L. Bonatti

plainly go through the list of this second class of arguments and show that either they are not conclusive or, when reasonable, they point at problems which are troublesome for the mental model theory as well. The arguments follow in no particular order of importance.

2. Mental logic doesn't have the machinery to deal with meaning and cannot explain the role of content and context in understanding and reasoning This is one of the major complaints against a mental logic. How could a formal theory enlighten us on such a clearly content-driven process as reasoning? In fact, as mental logic theorists recognize, one should distinguish two separate processes involved in problem solving. The first one is comprehension; the second one is reasoning proper. Accordingly, for mental logic theories a comprehension mechanism sensible to pragmatic information drives input analysis (Braine et al., 1984; Braine & O'Brien, 1991; O'Brien, 1993). Though the comprehension principles guiding it are only sketched, there is a hypothesis on their role in the time course of reasoning. After a first processing roughly delivering a syntactic analysis of a linguistic signal, the identification of its logical form and a first semantic analysis retrieving literal meaning, pragmatics and general knowledge aid to select a particular logical form for the input signal. Afterwards, representations possibly sharply different from the first semantic analysis are passed onto a processor blind to content and pragmatics. The general picture suggested, with some integration, looks like the diagram in Fig. 1. So a theory of mental logic cannot, and does not intend to, explain the role of content in reasoning, though it may help to locate how and when content and pragmatics interact with reasoning proper. From this point of view, the complaint is correct. However, models are no improvement; the thesis that "in contrast [to mental logic], the model theory has the machinery to deal with meaning" (Byrne, 1991, p. 77) is false. Models are supposed to be constructed either directly from perception, or indirectly from language. In the first case, no detailed account on how perception should generate models has been given.1 For linguistic models, a sketch of the procedures for their constructions exists. According to it, models are constructed from propositional representations via a set of procedures sometimes Sometimes it looks as if perceptual models in Marr's sense are considered to be equivalent to mental models in Johnson-Laird's sense (see Johnson-Laird, 1983a; Johnson-Laird et al., 1992, p. 421), but there are structural differences between the two constructs which make it difficult to accept the identification. To mention the most apparent one, perceptual models don't contain negation, but mental models do. For this reason, for each perceptual model there is an infinite number of mental models corresponding to it. A perceptual model of John scratching his head is a mental model of John scratching his head, but also of John not scratching his leg, of John not running the New York Marathon, of Mary being late for a date, and so on.

Reasoni theory

Comprehension mechanism

A first semantic analysis is elaborated; then, with the aid of pragmatic information and world knowledge, the logical form of the input sentence is selected

Rules ( basic proced + heurist strateg

I

Fallb strate

Figure 1. The place of pragmatics, comprehension mechanisms and reasoning prope in double squares.)

114

L. Bonatti

called procedural semantics. For example (Johnson-Laird & Byrne, 1991, p. 170 ff.), when given as input a sentence like The circle is on the right of the triangle a parser will start working and after some crunching the following information will be placed on the top of its stack: (The-circle-is . . .)-*Sentence ((1,0, 0)(A)(O)) The output of the parser is a couple containing both the grammatical description of the input ("sentence") and its semantical evaluation (in this case, an array containing numerical coordinates specifying the interpretation of the spatial relation, and the interpretations of the definite descriptions). Only at this point will procedural semantics take over and construct a model out of the propositional representation of the sentence; in this case, the model will be: A

O

that is, an image of a triangle to the left of the circle. Now, notice the following points. First, the procedures that construct models do not operate properly on natural language sentences, but on the logical forms of propositional representations. Thus procedural semantics presupposes logical forms. By the same token, procedural semantics presupposes the literal meaning of words and sentences, which have to be received as its input. As Johnson-Laird himself writes, "The reader should bear in mind that the present theory uses a procedural semantics to relate language, not to the world, but to mental models" (1983a, p. 248). Procedural semantics is essentially translation from mental representations to mental representations, not a function from mental representations to the world. But, then, if procedural semantics is not about literal meaning and logical forms, neither are mental models. Second, procedural semantics can work only if the output of the parser is not ambiguous: for example, scope relations must be already straightened out. The sentence (1) Every man loves a woman must be parsed to yield either (2) For all men x there is some woman y such that (x love y) or

Why should we abandon the mental logic hypothesis?

7/5

(3) For some woman y all men x are such that (x loves y) Only on the basis of one of them can procedural semantics yield a mental model. Thus the input to procedural semantics must be clear. Third, the possibility to construct the appropriate models of a text strictly depends on the expression power of the logical forms on which procedural semantics operates. To continue with the previous example, there are interpretations of (1) which don't correspond to either (2) or (3), involving a generic reading of the indefinite description. While it is not clear that a mental model can express the difference between a generic woman and a specific woman, this much is clear: // the logical form is not rich enough to articulate such a distinction, then mental models cannot represent it either, since they come from expressible logical forms. Thus the input to procedural semantics must be rich. Fourth, while the programs implementing the mental model theory described in Johnson-Laird (1983a) and Johnson-Laird et al. (1992) assume that the syntactic analysis of the input sentence plus word meaning is sufficient to determine its propositional content and logical forms, in a more natural setting the propositional content needed to construct the relevant mental models cannot be the first semantic analysis of the input, but the propositional content and the logical forms of the message it conveys. Now, by standard Gricean reasons the message conveyed in "Luca is a nice guy" in a text like Q: Is Luca a good philosopher? A: Well, let's say that Luca is a nice guy has something to do with my ability as a philosopher, and not with how much people like me. So if we take seriously the proposal that mental models are the kind of structure we build when comprehending a text, it is this contextual message that they must retain. A similar point can be made for metaphors, analogies, and all the cases in which the hearer/reader gathers information from an utterance aided by her general world knowledge, her understanding of relevance in communication, and other pragmatic factors. Now, since procedural semantics is proposed as a set of procedures extracting models from propositional representations, clearly the propositional representations on which it has to act in order to build the right mental models are not the results of a first semantic analysis of input sentences retrieving their literal meaning, but the analysis of their message in context, which, therefore, has to be retrieved before models are constructed. Procedural semantics works once all the disambiguations due to context, scope phenomena and retrieval of the speaker's intentions have taken place. To sum up, the input to procedural semantics presupposes both the literal meaning of the text and its logical form, and must be rich, clear, free from

116

L. Bonatti

Input

i

Parsing

T

Procedural semantics

First and second semantic analysis

Mental models

After a first semantic analysis is elaborated, pragmatic information and world knowledge aid to select the logical form of the input sentence Figure 2. The place of pragmatic and comprehension mechanisms in the mental model hypothesis.

structural ambiguities, and post-pragmatic. Thus when we begin to fill in the details, we come up with a sophisticated input analysis and we get the overall picture presented in Fig. 2, which for what concerns the role of pragmatics and meaning, has no difference from the mental logic picture - just as mental logic, procedural semantics and mental models presuppose, and do not explain, a theory of how pragmatics affects the selection of the correct message a set of utterances carries in the relevant situation. It could be objected that I am presenting a misleading picture, based on the algorithms implementing a small fraction of the mental model theory rather than on the theory itself. Algorithms are only a part of the story; with time, the rest will come. So Johnson-Laird et al. (1992) write: The process of constructing models of the premises is, in theory, informed by any relevant general knowledge, but we have not implemented this assumption, (p. 425)

But such "assumption" amounts to the solution to the frame problem, and the suspicion that it won't be implemented is more than warranted (Fodor, 1983). In any case, if the problem were solvable, it would still be the case that the retrieval of the relevant message would occur in the pre-modelic construction processes

Why should we abandon the mental logic hypothesis?

117

selecting the right logical forms and propositional contents which are input to procedural semantics. In fact, there is a litmus paper to test sensibility to content. The natural understanding of entailment seems to require a connection in content between antecedent and consequent. But the paradoxes of material implication allow false arbitrary antecedents to imply arbitrary consequents, regardless of their contents and even of their truth values. So if a theory of reasoning licenses them, it surely can't be advertised as the model to imitate for sensibility to content. Now, while in Braine and O'Brien's (1991) logical theory of implication the paradoxes are not available as theorems, mental models allow one to derive them as valid inferences (Johnson-Laird & Byrne, 1991). The reason is pretty clear. The mental model theory of connectives mainly consists of a variation on truth tables, and truth tables are only sensible to truth values, not to content connections or relevance. Thus besides their name, models have no advantage over mental logic to explain the role of content in reasoning, in any of the relevant senses of "content". They cannot explain literal meaning, nor meaning in situation, nor how pragmatics and general knowledge affect interpretation, and they don't seem to have the adequate structure to do it.

3. There is no mental logic because people make fallacious inferences People often reach conclusions which, if judged according to the canons of standard logic, are fallacious. And this should be a problem for a mental logic. The most glaring problem is that people make mistakes. They draw invalid conclusions, which should not occur if deduction is guided by a mental logic. (Johnson-Laird, 1983a, p. 25)

In less sophisticated versions, the argument notices that undergraduates make mistakes, and, worst of all, they show reiterate resistance to the teacher's efforts to correct them (Bechtel & Abrahansen, 1991, p. 168 ff.), or that they make more mistakes than what the average individual should innately know according to the logical competence mental logic attributes to people (Churchland, 1990, p. 283). In fact, mistakes come in different classes. They may be due to cognitive components not engaging reasoning proper, such as the comprehension stage or strategies of response selection; to performance failures; or to faulty competence. Any errors due to pre-deductive, comprehension mechanisms, or post-deductive, response selection strategies, can be accommodated by the two hypotheses roughly in the same way: the existence of such errors doesn't count against mental logic any more than it counts against mental models. Performance mistakes are explained away by mental models by indicating how models are built and handled by mechanisms non-proprietary of reasoning-mostly, mechanisms of working

118

L. Bonatti

memory storage and retrieval. A system based on mental logic can account for them in the same way. Errors of competence - as it were, directly generated by how the reasoning box is - are a more delicate matter. The question is to decide with respect to which point of reference they are errors. Does failure to apply excluded middle count as an error? Does the absence of reasoning schemata corresponding to material implication count? Classical logic - or, for that matter, any alternative logics cannot be a favored point of reference without further justifications. One major task of a psychological theory of deductive reasoning is to characterize what people take the right implications to be starting from certain premises, under ideal conditions. What could count as a systematic error in this context? Previous assumptions on the nature of rationality must be exploited. It can be argued, for example, that it is rational to proceed from truths to truths. On this basis, invalid reasoning processes could count as mistakes. If it could be shown that under ideal conditions people respond erratically to identical problems, or embody a rule which brings about a systematic loss of truths, then it may be said that subjects make mistakes in point of competence regardless of the compliance of natural logical consequence to classical, or other, logics. But if this were the case, mental models would be in a worse position than mental logic. It is possible (though not desirable) to account for systematic errors within a mental logic framework by indicating which rules (if any) induce systematic violations of the selected normative model. As of today the algorithms proposed to implement logical reasoning by models are either psychologically useless or ill defined (Bonatti, in press; O'Brien, Braine & Yang, in press), so it is difficult to give a definite judgement on this issue, but the tentative set of rules proposed for model construction is meant to be truth preserving in principle. Thus it is puzzling to figure out how models might account for purported systematic violations: errors in point of competence would be an even deeper mystery for the mental model hypothesis.

4. There is no mental logic because higher-order quantifiers are not representable in first-order logic, and yet we reason with them This argument has been considered "the final and decisive blow" to the doctrine of mental logic (Johnson-Laird, 1983a, p. 141). According to Barwise and Cooper (1981), expressions such as "More than half of" or "Most" are sets of sets, and therefore an adequate logic for natural language needs to extend beyond first order. The argument from this proposal to the rejection of mental logic runs as follows: [Higher-order calculus] is not complete. If there can be no formal logic that captures all the valid

Why should we abandon the mental logic hypothesis?

119

deductions, then a fortiori there can be no mental logic that does either. It is a remarkable fact that natural language contains terms with an implicit "logic" that is so powerful that it cannot be completely encompassed by formal rules of inference. It follows, of course, that any theory that assumes that the logical properties of expressions derive directly from a mental logic cannot give an adequate account of those that call for a higher-order predicate calculus. This failure is a final and decisive blow to the doctrine of mental logic. (Johnson-Laird, 1983a, pp. 140-141, italics mine)

The argument has often been repeated (see, for example, Johnson-Laird & Bara, 1984, p. 6; Johnson-Laird & Byrne, 1990, p. 81; Johnson-Laird & Byrne, 1991, p. 15); so it must be attached a certain importance. The question is to figure out why. The nature of the representational device in which mental processes are carried out is an empirical question, and if patterns of inference are required that can be better formalized in second-order logic, so be it. So what can possibly be wrong in using higher-order logic? We are told, it is not complete. Such objection makes sense only if one presupposes that a mental calculus must be complete. But an argument is needed to ask for completeness as a constraint over a mental logic, and it is difficult to see what it would look like. We may impose constraints on a logical system by requiring that it possesses certain logical properties such as consistency, or completeness, because we can decide what we want from it. But finding out how people reason is an empirical enterprise. It would be a very interesting empirical discovery to find out that, say, a subject's system for propositional reasoning is complete, but it's not enough that we want it to be so. Even more basic logical properties cannot be granted a priori. It would be desirable that subjects reason consistently, as everybody hopes to discover that under ideal conditions they do, but, again, to presuppose that our reasoning system is consistent requires an argument. Barring such arguments, the "final and decisive blow against mental logic" blows up. In fact, it may backfire. Johnson-Laird et al. blame the incompleteness of a higher-order mental logic system as if the mental model counterproposal were complete. But the only fragment for which a psychological implementation has been proposed - propositional reasoning-is not even valid. Models have no advantage over mental logic on the issue of completeness. Neither should they: such an advantage, in the absence of evidence that natural reasoning is complete, would be irrelevant.

5. There is no evolutionary explanation of the origin of mental logic Another alleged argument against mental logic concerns its origin. A bland version of it simply claims that there is no evolutionary explanation of mental logic, and this is enough to reject the theory (Cosmides, 1989). A richer version runs as follows. To accept that there is a mental logic seems to lead to the

120

L. Bonatti

admission that most of our reasoning abilities are innate. Nativism, in general, cannot be a problem: everybody has to live with it, and the only issue is whether you like it weaker or stronger. But there should be something specifically wrong with nativism about mental logic: there is no evolutionary explanation for its origin: By default, it seems that our logical apparatus must be inborn, though there is no account of how it could have become innately determined (Johnson-Laird, 1983a, p. 40). The moral that Fodor drew is an extreme version of nativism - no concept is invented; all concepts are innate. Alas, any argument that purports to explain the origins of all intellectual abilities by postulating that they are innate merely replaces one problem by another. No one knows how deductive competence could have evolved according to the principles of neo-Darwinism. (JohnsonLaird, 1983a, pp. 142-143) So intractable is the problem for formal rules that many theorists suppose that deductive ability is not learned at all. It is innate. Fodor (1980) has even argued that, in principle, logic could not be learned. The difficulty with this argument is not that it is wrong, although it may be, but that it is too strong. It is hard to construct a case against the learning of logic that is not also a case against its evolution. If it could not be acquired by trial-and-error and reinforcement, then how could it be acquired by neo-Darwinian mechanisms? (Johnson-Laird & Byrne, 1991, p. 204)

It is first worth noticing that the argument is meant to apply to cognition, and only to very restricted kinds of cognitive abilities. If you try to generalize it beyond this domain, it becomes flatly absurd. For the given premise is that Darwinian mechanisms are a sort of trial-and-error and reinforcement mechanisms applied to the species. Its generalization says: for any *, if x cannot be acquired by trial-and-error and reinforcement, then how could it be acquired by a neo-Darwinian mechanism? Now take a non-cognitive phenomenon and substitute it for x; breathing cannot be acquired by trial-and-error and reinforcement, so how did the species acquire the ability to breathe? That doesn't work. And neither does it work for most innate cognitive abilities. Try with colors, or perceptual primitives: the ability to recognize colors (or any perceptual primitive) cannot be acquired by trial-and-error and reinforcement, so how could the ability to recognize colors be acquired by neo-Darwinian mechanisms? This doesn't work either. So I assume that the argument is really targeted against mental logic. Second, even restricting its field of application, notice that there are at least three different questions one may raise. What is the logical syntax of mental processes? What logical system underlies reasoning abilities? What concepts is the mind able to entertain, whether innately or by experience? The above argument does not keep them separate, yet they may have radically different answers. For example, an organism may be innately endowed with the syntax of first-order logic, but it may keep changing its logical system (for simplicity, the set of its axioms) by flip-flopping an axiom, and at the same time may need to learn any concept by experience. Such an organism would have an innate logical syntax, but no innate logic or innate concepts. Or else, an organism may be endowed with an

Why should we abandon the mental logic hypothesis?

121

innate logical syntax and an innate logic, but may need experience to acquire contentful concepts. The arguments for or against nativism are quite different in the three cases. I will assume that the above argument is really targeted against nativism of a system of logic. Then, it can be reconstructed in the following way. If there is a mental logic, an account is due of how it is acquired. Since there is no theory of its acquisition, it must be assumed that the logical system - not just its syntax - is innate. But, alas, this claim is unsupported because there is no evolutionary story on how such a system gets fixated. Thus the doctrine of mental logic has to be rejected. The short answer to such an argument (in its bland and its rich forms) is: too bad for evolutionary explanations. The long answer requires a reflection on the state of evolutionary explanations of cognitive mechanisms. The argument presupposes that there must be an evolutionary explanation of how deductive abilities are fixated. What would it look like? For the much clearer case of language, evolutionary explanations are uninformative. Whether a mutation endowing humans with linguistic abilities concerns the structures of the organism or in its functions; whether language has been a direct mutation, or a byproduct of another mutation; under what metric it turned out to be advantageous: these are unanswered questions. This is a general problem concerning the application of evolutionary concepts to cognition. The quest for a Darwinian explanation of cognitive evolution is founded at best on an analogy with biological evolution, and analogies may be misleading. Lewontin specifically makes this point for problem solving: . .. generalized problem solving and linguistic competence might seem obviously to give a selective advantage to their possessors. But there are several difficulties, First, . . . human cognition may have developed as the purely epiphenomenal consequence of the major increase in brain size, which, in turn, may have been selected for quite other reasons. .. . Second, even if it were true that selection operated directly on cognition, we have no way of measuring the actual reproductive advantages. . . . Fourth, the claim that greater rationality and linguistic ability lead to greater offspring production is largely a modern prejudice, culture - and history - bound. . . . The problem is that we do not know and never will. We should not confuse plausible stories with demonstrated truth. There is no end to plausible story telling. (Lewontin, 1990, pp. 244-245)

And there is no reason to ask for mental logic what does not exist and might not exist for other, better-known, cognitive domains. But let us suppose that one should seriously worry for the lack of a Darwinian explanation of how innate logic has been selected. Again, here one should sense the kind of comparative advantage that the mental model hypothesis gains. The argument seems to presuppose that, as opposed to the case of mental logic, either (a) the ability of building mental models is not innate but learned, and thus Darwinian worries don't arise, or (b) if it is not learned, there is an evolutionist explanation of its origin. Alternative (a) is empty. There is no learning theory for models and it is

122

L. Bonatti

unlikely that any future such theory will bring about substantial economies in nativism, since most of the structures needed for problem solving are the same regardless of which theory turns out to be correct. Without (a), alternative (b) assumes the following form: an innate mechanism for building mental models gives an evolutionary advantage that an innate mental logic doesn't give. But evolutionary explanations are not so fine-grained to discriminate between our capacity to construct models as opposed to derivations. If there is any such explanation, it will work for both; if there isn't one for mental logic, there isn't one for mental models either.

6. Mental logic cannot explain reasoning because people follow extra-logical heuristics Often heuristics of various sorts guide human responses even in deduction. But it is unclear how this counts against mental logic. Models need heuristics as much as logical rules do. For example, if a premise has different possible interpretations, an order is needed to constrain the sequence of constructed models (Galotti, 1989). Such an order too may depend on heuristics having nothing to do with models proper, such as reliance on the most frequent interpretation, or on previous experience, or on previously held beliefs. But there may be something more to the argument. It may be argued that heuristics don't pose any special problem to model-based theories of reasoning, whereas they do for logic-based theories. Just like Dennett's queen moving out early, heuristics can be an epiphenomenon of the structure of models, whereas rule-based systems must express them explicitly. For example, a model for the sentences "a is to the right of &" and "6 is to the right of c" allows us to derive "a is to the right of c" with no explicit rule to that effect (see Johnson-Laird, 1983a; Johnson-Laird & Byrne, 1991). In this case, transitivity is an emerging feature of the structure of the model. Analogously, it may be argued that also other apparent rule-following behaviors such as strategies are emerging features of models. However, often subjects reason by following heuristics that they can perfectly spell out and that are not accounted for by the structure of models (see, for example, Galotti, Baron & Sabini, 1986), and this squares very badly with a radical rule epiphenomenalism. At least in principle, models may help to solve the problem of implicitness: certain processes may be externally described by explicit rules which nevertheless are not explicitly represented in the mental life of an organism. Solution: the rules supervene to the structure of models. But the other side of the coin is the problem of explicitness: how could a system represent the information that is explicitly represented? This is no difficulty for mental logic, but how could a heuristic be explicitly represented within models? Tokens and possibly some of their logical relations are explicit in models, but not

Why should we abandon the mental logic hypothesis?

123

performatives. Models don't contain information specifying the order in which certain operations have to be executed, but only the result of such operations. So while a propositional-like system doesn't have the problem of explicitness, models may have it. 7. Mental logic cannot offer a theory of meaning for connectives In fact, the formal rules for propositional connectives are consistent with more than one possible semantics . .. Hence, although it is sometimes suggested that the meaning of a term derives from, implicitly reflects, or is nothing more than the roles of inference for it, this idea is unworkable . . . (Johnson-Laird et al., 1992, p. 420)

But truth tables (and thus models) don't have such a problem, since they "are merely a systematic way of spelling out a knowledge of the meanings of connectives" (Johnson-Laird et al., 1992, p. 420). Johnson-Laird et al. refer to an argument presented in Prior (1960, 1964). But an aspect of it has been forgotten. Prior argued that rules of inference cannot analytically define the meaning of the connectives they govern. If there were nothing more to the meaning of a connective than the inferences associated to it, then the connective tonk could be defined, with the meaning specified by the following rules: (1) From P, derive P tonk Q (2) From P tonk Q, derive Q and with tonk we could obtain the following derivation: 2 and 2 are 4 Therefore, 2 and 2 are 4 tonk 2 and 2 are 5 Therefore, 2 and 2 are 5. Prior's argument is a challenge to a conceptual role semantics. If meaning is inferential role, how to avoid tonkl According to Prior, tonk shows that explicit definitions cannot give the meaning to a term on the ground of the analytical tie between the definiens and the definiendum, but can at most correspond to a previously possessed meaning: we see that certain rules of inferences are adequate for "and" because we know its meaning and judge the adequacy of the rules with respect to it. We can perfectly introduce a sign for tonk governed by the above rules and have a purely symbolic game running. But games with rules and transformations of symbols don't generate meaning: "to believe that anything of this sort can take us beyond the symbols to their meaning, is to believe in magic" (Prior, 1964, p. 191). The difference between "and" and tonk is that in the first case the rules correspond to the (previously held) sense of the word "and": they

124

L. Bonatti

don't confer it its meaning, but are "indirect and informal ways" (Prior, 1964, p. 192) to clarify it. But in the second case there is no prior sense to appeal to. We can define a class of signs standing for conjunction, and a class of signs standing for contonktion, but the latter is an empty class. There are conjunction-forming signs, because there is a conjunction. There are no contonktion-forming signs, because there is no contonktion and the explicit introduction of a sign for it does not give life to a new connective. So Prior's argument goes. One way to read it is that rules can't give a symbol its meaning, but something else can: namely, truth tables in a metalanguage. This seems to be the interpretation adopted by Johnson-Laird et al. (1992) when they claim that mental logic cannot explain the meaning of connectives, but truth tables can. In fact, Prior (1964) remarked that explicitly defining connectives in terms of truth tables did not change the point of his criticism. In his view, there was "no difference in principle between [rules of inferences and truth tables]" (Prior, 1964, p. 192). Instead of using rules, he argued, we can define a conjunctionforming sign by using the familiar truth table, but this will not give conjunction its meaning; any formula of arbitrary length with the same truth table will turn out to be a conjunction-forming sign; so will formulas involving non-logical conceptions such as "P ett Q", which is the abbreviation for "Either P and Q, or Oxford is the capital of Scotland" (Prior, 1964, p. 194). The point of this further facet of the argument is that truth tables identify a much broader class of signs than conjunction, and moreover, signs that are understood on the basis of the understanding of conjunction (see Usberti, 1991). We might try to eliminate all the unwanted signs which would be defined by the truth table for conjunction by saying that the table defines the meaning of the shortest possible sign for conjunction. We would probably be happy with this solution. But, Prior noticed, we would accept it because we understand that such a characterization captures the meaning of the conjunction, and not of the other signs. Thus, truth tables are in no better position than rules to generate meanings. If they apparently don't suffer from tonkitis, they suffer from another equally worrisome disease. And if we wanted to resort again to formal games, then tonkitis would reappear, since a (symbolic) truth table game defining a contonktion-forming sign is easy to find: tonk "is a sign such that when placed between two signs for the propositions P and Q, it forms a sign that is true if P is true and false if Q is false (and therefore, of course, both true and false if P is true and Q is false)" (Prior, 1964, p. 193). We can now leave Prior and touch on the real problem. If we grant that explicit rules, or truth tables, don't define the meaning of the logical symbols, but are accepted on the basis of their correspondence to some pre-existent meaning we

Why should we abandon the mental logic hypothesis?

125

attach to connectives and quantifiers, we still have to explain what the source of our intuitions about the meaning of connectives and quantifiers is, because if thinking that a game of symbols can take us beyond the symbols to their meanings is magic, as Prior said, it is equally magic to think that the meaning of logical symbols comes from nowhere. For Johnson-Laird, Byrne, and Schaeken, truth tables are merely "a systematic way of spelling out a knowledge of the meanings of connectives". But in general this is false. There are 16 binary truth tables: only some of them do, or seem to, spell out the meaning of binary connectives; others clearly don't. Why is it so? Why do we feel that the truth table, for the conjunction reflects the meaning of the conjunction, whereas the classical truth table for the implication doesn't reflect the meaning of natural implication, and the anomalous truth table for tonk can't reflect the meaning of a new connective? Nothing seems to block the following possibility. When I see somebody who reminds me of my brother, one of the possibilities is that it is my brother. So when I see a set of rules for the conjunction and I think that it adequately expresses what I mean by a conjunction, one of the possibilities is that I find that resemblance because the rules are the exact expression of the patterns of inferences of a logical connective in the mind. In this case, there is nothing more to the meaning of the term than the rules themselves. At the same time, when I see the truth table of material implication I realize that it does not spell out the meaning of natural implication because the rules governing natural implications are not reflected in it, and when I see the rules of inference-or the truth table-for tonk, I have no intuition about their adequacy because there is no logical connective for tonk in the mind, from which the explicit rules are a clone copy. Contonktion cured. Intuitions, however, are not good guides. It is not enough to say that conjunctions have a meaning because they seem to correspond to rules in the mind but contonktions don't because they don't titillate our intuitions. There are lots of logical operators that may not have any straightforward correspondence with natural language, and yet are computed in retrieving the truth conditions of natural language sentences - consider, for example, focus, or quantifiers over events. If a semanticist presented us with a set of rules for them, we would not probably have the same immediate intuition we feel for conjunction. This is where a theory of mental logic comes in. A developed theory of mental logic offers empirical reasons to show that conjunctions are in the mind, while contonktions are not. If such a theory can be worked out (and a tiny part of it already exists), then mental logic can be the basis of a theory of meaning for natural connectives. For the moment, we are very far from having such a complete theory. The present point is simply that no argument exists to hamper its development.

126

L. Bonatti

8. There is no mental logic because valid inferences can be suppressed This recent argument is based on the so called "suppression of valid inferences" paradigm. By modifying a paradigm used by Rumain, Connell, and Braine (1983), Byrne (1989) set up an experiment in which to premises such as If she meets her friend she will go to a play She meets her friend an extra premise was added, transforming the argument in If she meets her friend she will go to a play If she has enough money she will go to a play She meets her friend and she showed that in this case the percentage of subjects applying modus ponens drops from 96% to 38%. Mental model theorists attributed a considerable importance to this result. It shows, they claimed, that also valid deductions as strong as modus ponens can be blocked: Models can be interrelated by a common referent or by general knowledge. Byrne (1989) demonstrated that these relations in turn can block modus ponens. . .. The suppression of the deduction shows that people do not have a secure intuition that modus ponens applies equally to any content. Yet, this intuition is a criterion for the existence of formal rules in the mind. (Johnson-Laird et al, 1992, p. 326)

and as a consequence that by their own argument, rule theorists ought to claim that there cannot be inference rules for (valid deduction). (Johnson-Laird & Byrne, 1991, p. 83)

But no argument is offered to ensure that modus ponens is really violated, or to justify the claim that this result supports the mental models hypothesis. If we assume that deductive rules apply not to the surface form of a text, but to its integrated representation, then subjects may be led by pragmatic reasons to construe the two premises If she meets her friend she will go to a play If she has enough money she will go to a play as a single If (she meets her friend and she has enough money) she will go to a play

Why should we abandon the mental logic hypothesis?

127

and therefore when provided only with the premise "She meets her friend", they don't know about the truth of the conjunctive antecedent and correctly refuse to use modus ponens. In other cases, also studied by Byrne, when subjects are given arguments such as If she meets her friend she will go to a play If she meets her mother she will go to a play She meets her friend they do conclude that she will go to a play. This may be because subjects compose the premises "If A then B" and "If C then B" as a single "If (A or C), then B", and knowing one of the disjuncts of the composed antecedent suffices to correctly apply modus ponens. Thus, under this interpretation, there is no suppression of valid inferences: simply, people tend to construct a unified representation of a text which may itself be governed by formal rules of composition. It may be replied that my response to the suppression argument puts the weight of the explanation on pre-logical comprehension processes, rather than on deduction proper, and that mental logic theorists have no account of such processes. This wouldn't be necessary for models, because they "have the machinery to deal with meaning". But I have shown that such a claim is false. Models too rely on pragmatic comprehension mechanisms, and don't explain them. If model theorists want to explain why people draw the inference in one case and not in the other, they have to say that in one case a model licensing the inference is constructed, and in the other a model not licensing the inference is constructed. To account for why it is so, they offer no explanation.

9. Conclusions "Yes, but mental logic has had its shot. It has been around for centuries and nothing good came out of it. It's time to change." Often the contrast between the long history of mental logic and its scarce psychological productivity is taken as a proof of its sterility. In fact, this impression derives from a mistake of historical perspective. The idea is very ancient, but the conceptual tools needed to transform it into the basis for testable empirical hypotheses are very recent. For centuries, logic too remained substantially unchanged, to the point that Kant considered it a completed discipline (1965, pp. 17-18). So there was no reason to change the conventional wisdom on the relations between logic and psychology: the former was stable because considered complete and the latter was stable because non-existent. When, with Frege, Russell and the neopositivists, logic as we mean it started being developed, the routes of logic and psychology separated.

128

L. Bonatti

Well beyond the 1930s, among the large majority of philosophically minded logicians, showing interest in psychological processes became a sort of behavior that well-mannered people should avoid. No substantial argument against the psychological feasibility of mental logic motivated this change of view. Rather, its roots have to be looked for in the general spirit of rebellion against German and English idealism from which twentieth-century analytic philosophy stemmed. Nevertheless, for independent reasons, the same conclusion became popular among experimental psychologists and was generally held until the early 1960s, both by behaviorists and by the new-look psychologists. There was, indeed, the Piagetian exception, but it does not count: Piaget's flirting with mental logic was never clear enough to become a serious empirical program (Braine & Rumain, 1983), and recent Piagetian-oriented investigations on mental logic (see Overton, 1990) have not helped towards a clarification. It was again an impulse coming from logicians - not from psychologists - that put logic back in the psychological ballpark. Hilbert first directly expressed a connection between symbols and thought which could serve as a psychological underpinning for mental logic. For him, the fundamental idea of proof theory was "none other than to describe the activity of our understanding, to make a protocol of the rules according to which our thinking actually proceeds. Thinking, it so happens, parallels speaking and writing: we form statements and place them one behind another" (1927, p. 475). Yet Hilbert's intuition was not enough. Formal systems, as conceived by the axiomatic school, were the least possible attractive tool to investigate the psychology of reasoning. What was still missing to render logic ready for psychological investigation was on the one side a more intuitive presentation of formal systems, and on the other side a model of how a physical structure can use a formal system to carry out derivations. The first was provided by Gentzen, and the second by Turing. However, once again, the distance between Gentzen's and Turing's ideas and a real psychological program should not be underestimated. Gentzen did introduce the systems of natural deduction with the aim to "set up a formal system which comes as close as possible to actual reasoning" (Gentzen, 1969, p. 68), but his reference to "actual reasoning" was merely intuitive. And Turing did offer the abstract model of how a physical mechanism could perform operations once considered mental along the lines suggested by Hilbert, but Turing's real breakthrough consisted of the realization that a computer can be a mind, namely, that certain kinds of properties once attributable only to humans can also be appropriately predicated of other physical configurations. Such insight, however, leaves the mechanisms and procedures by which the mind itself operates underspecified. It says that mental processes can be simulated, but it leaves it undetermined whether the simulandum and the simulans share the same psychology. The further step necessary to the formulation of a psychological notion of mental logic came when functionalism advanced the explicit thesis that the

Why should we abandon the mental logic hypothesis?

129

psychological vocabulary is computational vocabulary, and that the natural kinds described by psychology are not organisms, but computational devices. The change leading to this second step was gradual, and required a lot of philosophical work to be digested. We are now beyond the 1960s, and not in Aristotle's age. Only then had logic and philosophy come to the right point of development to take the mental logic hypothesis seriously. And another decade or more had to go before experimental techniques were sufficiently developed to begin asking nature the right questions in the right way. The works by Braine, Rips and their collaborators are the first attempts at elaborating mental logic in a regimented psychological setting. Thus the psychological history of mental logic is very recent. It is, in fact, roughly contemporary with the psychological history of the mental model hypothesis. This shouldn't come as a surprise: both needed largely the same conceptual tools to be conceived. Mental models are not the inevitable revolution after millennia of mental logic domination. So, contrary to widespread assumptions, there are no good arguments against mental logic, be it point of principle, or in point of history. If a case against it and in favor of mental models can be made, it cannot rest on principled reasons, but on the formal and empirical development of the two theories. Indeed, extending the mental logic hypothesis beyond propositional reasoning engenders formidable problems connected with the choice of an appropriate language to express the logical forms of sentences on which rules apply, the choice of psychologically plausible rules to test, and the choice of appropriate means to test them. Approaching these problems requires the close collaboration of psychologists, natural language semanticists and syntacticians. But these are problems, however hard, and not mysteries. Most psychologists have abandoned the program and married the mental models alternative, both for its supposed superiority in handling empirical data and for the overwhelmingly convincing arguments against mental logic. In fact, the case for mental models has been overstated under both counts. Given how little we know about the mind and reasoning, conclusions on research programs that only began to be adequately developed a few years ago are premature. Psychologists should keep playing the mental logic game.

References Barwise, J., & Cooper, R. (1981). Generalized quantifiers and natural language. Linguistics and Philosophy, 4, 159-219. Bechtel, W., & Abrahansen, A. (1991). Connectionism and the mind. Oxford: Basil Blackwell. Bonatti, L. (in press). Propositional reasoning by model? Psychological Review. Boolos, G. (1984). On 'Syllogistic inference'. Cognition, 17 181-182. Braine, M.D. (1979). If then and strict implication; A response to Grandy's note. Psychological Review, 86, 158-160.

130

L. Bonatti

Braine, M.D., & O'Brien, D.P. (1991). A theory of if: Lexical entry, reasoning program, and pragmatic principles. Psychological Review, 98, 182-203. Braine, M.D., Reiser, B.J., & Rumain, B. (1984). Some empirical justification for a theory of natural propositional logic. The psychology of learning and motivation (Vol. 18, pp. 313-371). San Diego, CA: Academic Press. Braine, M.D., & Rumain. B. (1983). Logical Reasoning. In J. Ravell (Ed.), CaarmichaeVs handbook of child psychology (Vol. Ill, pp. 263-340). New York: Wiley. Byrne, R. (1989). Suppressing valid inferences with conditionals. Cognition, 31, 61-83. Byrne, R. (1991). Can valid inferences be suppressed? Cognition, 39, 71-78. Churchland, P.M. (1990). On the nature of explanation: a PDP approach, reprinted in S. Forrest (1991), Emergent computation (pp. 281-292). Cambridge, MA: MIT Press. Cosmides, L. (1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason selection task. Cognition, 31, 187-276. Ehrlich, K., & Johnson-Laird, P.N. (1982). Spatial descriptions and referential continuity. Journal of Verbal Learning and Verbal Behavior, 21, 296-306. Fodor, J.A. (1980). On the impossibility of acquiring 'more powerful' structures. In M. PiatteliPalmarini (Ed.), Language and learning: The debate between Jean Piaget and Noam Chomsky (pp. 142-163). Cambridge MA: Harvard University Press. Fodor, J.A. (1983). Modularity of mind. Cambridge, MA: MIT Press. Galotti, K.M. (1989). Approaches to studying formal and everyday reasoning. Psychological Bulletin, 105, 331-351. Galotti, K.M., Baron, J., & Sabini, J.P. (1986). Individual differences in syllogistic reasoning: deduction, rules or mental models? Journal of Experimental Psychology. General, 115, 16-25. Garnham, A. (1987). Mental models as representations of discourse and text. Chichester: Ellis Horwood Ltd. Gentzen, G. (1969). Investigations into logical deduction. In M.E. Szabo (Ed.), The collected papers of Gehrard Gentzen (pp. 68-128). Amsterdam: North Holland. Hilbert, D. (1927). The foundations of mathematics. In J. van Heijenoort (Ed.), From Frege to Godel (pp. 464-469). Cambridge, MA: Harvard University Press 1967. Hodges, J. (1993). The logical content of theories of deduction. Behavioral and Brain Sciences, 16 353-354. Johnson-Laird, P.N. (1983a). Mental models. Cambridge, MA: Harvard University Press. Johnson-Laird, P.N. (1983b). Thinking as a skill. In J.B. Evans (Ed.), Thinking and reasoning: psychological approaches (pp. 44-75). London: Routledge & Keegan. Johnson-Laird, P.N. (1989). Mental models. In M. Posner (Ed.), Foundations of cognitive science (pp. 469-499). Cambridge, MA: MIT Press. Johnson-Laird, P.N., & Bara, B. (1984). Syllogistic inference. Cognition, 16, 1-61. Johnson-Laird, P.N., & Byrne, R. (1989). Spatial reasoning. Journal of Memory and Language, 28, 565-575. Johnson-Laird, P.N., & Byrne, R. (1990). Meta-logical problems: Knights, knaves and Rips, Cognition, 36, 69-84. Johnson-Laird, P.N., & Byrne, R. (1991). Deduction. Hillsdale, NJ: Erlbaum. Johnson-Laird, P.N., & Byrne, R.M. (1993). Precis of Deduction. Behavioral and Brain Sciences, 16, 323-380. Johnson-Laird, P.N., Byrne, R., & Schaeken, W. (1992) Propositional reasoning by model. Psychological Review, 99, 418-439. Johnson-Laird, P.N., Byrne, R.M., & Tabossi, P. (1989). Reasoning by model: the case of multiple quantification. Psychological Review, 96, 658-673. Kant, I. (1965). Critique of pure reason. New York: St. Martin Press. Lea, R.B., O'Brien, D.P., Fisch, S., Noveck, I., & Braine, M. (1990). Predicting propositional logic inferences in text comprehension. Journal of Memory and Language, 29, 361-387. Lewontin, R.C. (1990). The evolution of cognition. In D. Osherson and E. Smith (Eds.), Thinking: an invitation to cognitive science (vol. 3, pp. 229-246). Cambridge, MA: MIT Press. McGinn, C. (1989). Mental content. Oxford: Basil Blackwell.

Why should we abandon the mental logic hypothesis?

131

Oakill, J., Johnson-Laird, P.N., & Garnham, A. (1989). Believability and syllogistic reasoning. Cognition, 31, 117-140. O'Brien, D. (1993). Mental logic and irrationality: we can put a man on the moon, so why can't we solve those logical reasoning problems? In K.I. Manktelow & D.E. Over (Eds.), Rationality (pp. 110-135). London: Routledge. O'Brien, D., Braine, M.D., & Yang, Y. (in press). Proportional reasoning by mental models? Simple to refute in principle and in practice. Psychological Review. Overton, W. (Ed.) (1990). Reasoning, necessity and logic: Developmental perspectives. Hillsdale, NJ. Erlbaum. Prior, A.N. (1960). The runabout inference ticket. Analysis, 21, 38-39. Prior, A.N. (1964). Conjunction and contonktion revisited. Analysis, 24, 191-195. Rips, L.J. (1983). Cognitive processes in propositional reasoning. Psychological Review, 90, 38-71. Rips, L.J. (1986). Mental muddles. In M. Brand & R. Harnish (Eds.), The representation of knowledge and belief (pp. 258-286). Tucson: University of Arizona Press. Rumain, B., Connell, J., & Braine, M.D. (1983). Conversational comprehension processes are responsible for reasoning fallacies in children as well as adults: It is not the biconditional. Developmental Psychology, 19, 471-481. Usberti, G. (1991). Prior's disease, Teoria, 2, 131-138.

7

Concepts: a potboiler Jerry Fodor* Graduate Center, CUNY, 33 West 42nd Street, New York, NY 10036, USA Center for Cognitive Science, Rutgers University, Psychology Building, Busch Campus, Piscataway, NJ 08855, USA

Abstract An informal, but revisionist, discussion of the role that the concept of a concept plays in recent theories of the cognitive mind. It is argued that the practically universal assumption that concepts are (at least partially) individuated by their roles in inferences is probably mistaken. A revival of conceptual atomism appears to be the indicated alternative.

Introduction: the centrality of concepts What's ubiquitous goes unremarked; nobody listens to the music of the spheres (or to me, for that matter). I think a certain account of concepts is ubiquitous in recent discussions about minds; not just in philosophy but also in psychology, linguistics, artificial intelligence, and the rest of the cognitive sciences; and not just this week, but for the last fifty years or so. And I think this ubiquitous theory is quite probably untrue. This paper aims at consciousness raising; I want to get you to see that there is this ubiquitous theory and that, very likely, you yourself are among its adherents. What to do about the theory's not being true (if it's not) - what our cognitive science would be like if we were to throw the theory overboard-is a long, hard question, and one that I'll mostly leave for another time. The nature of concepts is the pivotal theoretical issue in cognitive science; it's the one that all the others turn on. Here's why: Cognitive science is fundamentally concerned with a certain mind-world

* Correspondence to: J. Fodor, Center for Cognitive Science, Rutgers University, Psychology Building, Busch Campus, Piscataway, NJ 08855, USA.

J. Fodor

134

relation; the goal is to understand how its mental processes can cause a creature to behave in ways which, in normal circumstances, reliably comport with its utilities. There is, at present, almost1 universal agreement that theories of this relation must posit mental states some of whose properties are representational, and some of whose properties are causal. The representational (or, as I'll often say, semantic) properties of a creature's mental states are supposed to be sensitive to, and hence to carry information about, the character of its environment.2 The causal properties of a creature's mental states are supposed to determine the course of its mental processes, and, eventually, the character of its behavior. Mental entities that exhibit both semantic and causal properties are generically called "mental representations", and theories that propose to account for the adaptivity of behavior by reference to the semantic and causal properties of mental representations are called "representational theories of the mind". Enter concepts. Concepts are the least complex mental entities that exhibit both representational and causal properties; all the others (including, particularly, beliefs, desires and the rest of the "propositional attitudes") are assumed to be complexes whose constituents are concepts, and whose representational and causal properties are determined, wholly or in part, by those of the concepts they're constructed from. This account subsumes even the connectionist tradition which is, however, often unclear, or confused, or both about whether and in what sense it is committed to complex mental representations. There is a substantial literature on this issue, provoked by Fodor and Pylyshyn (1988). See, for example, Smolensky (1988) and Fodor and McLaughlin (1990). Suffice it for present purpose that connectionists clearly assume that there are elementary mental representations (typically labeled nodes), and that these have both semantic and causal properties. Roughly, the semantic properties of a node in a network are specified by the node's label, and its causal properties are determined by the character of its connectivity. So even connectionists think there are concepts as the present discussion understands that notion. On all hands, then, concepts serve both as the domains over which the most elementary mental processes are defined, and as the most primitive bearers of semantic properties. Hence their centrality in representational theories of mind.

l

The caveat is because it's moot how one should understand the relation between main-line cognitive science and the Gibsonian tradition. For discussion, see Fodor and Pylyshyn (1981). 2 There is no general agreement, either in cognitive science or in philosophy, about how the representational/semantic properties of mental states are to be analyzed; they are, in general, simply taken for granted by psychologists when empirical theories of cognitive processes are proposed. This paper will not be concerned, other than tangentially, with these issues in the metaphysical foundations of semantics. For recent discussion, however, see Fodor (1990) and references cited therein.

Concepts: a potboiler

135

1. Ancient history: the classical background The kind of concept-centered psychological theory I've just been sketching should seem familiar, not only from current work in cognitive science, but also from the philosophical tradition of classical British empiricism. I want to say a bit about classical versions of the representational theory of mind because, though their general architecture conforms quite closely to what I've just outlined, the account of concepts that they offered differs, in striking ways, from the ones that are now fashionable. Comparison illuminates both the classical and the current kinds of representational theories, and reveals important respects in which their story was closer to being right about the nature of concepts than ours. So, anyhow, I am going to argue. Here's a stripped-down version of a classical representational theory of concepts. Concepts are mental images. They get their causal powers from their associative relations to one another, and they get their semantic properties from their resemblance to things in the world. So, for example, the concept DOG applies to dogs because dogs are what (tokens of) the concept looks like. Thinking about dogs often makes one think about cats because dogs and cats often turn up together in experience, and it's the patterns in one's experience, and only these, that determine the associations among one's ideas. Because association is the only causal power that ideas have, and because association is determined only by experience, any idea can, in principle, become associated to any other, depending on which experiences one happens to have. Classical ideas cannot, therefore, be defined by their relations to one another. Though DOGthoughts call up CAT-thoughts, LEASH-thoughts, BONE-thoughts, BARKthoughts and the like in most actual mental lives, there are possible mental lives in which that very same concept reliably calls up, as it might be, PRIME NUMBER-thoughts or TUESDAY AFTERNOON-thoughts or KETCHUPthoughts. It depends entirely on how often you've come across prime numbers of dogs covered with ketchup on Tuesday afternoons. So much by way of a reminder of what classical theorists said about concepts. I don't want to claim much for the historical accuracy of my exegesis (though it may be that Hume held a view within hailing distance of the one I've sketched; for purposes of exposition, I'll assume he did). But I do want to call your attention to a certain point about the tactics of this kind of theory construction - a point that's essential but easy to overlook. Generally speaking, if you know what an X is, then you also know what it is to have an X. And ditto the other way around. No doubt, this applies to concepts. If, for example, your theory is that concepts are pumpkins, then it has to be a part of your theory that having a concept is having a pumpkin; and if your theory is that having a concept is having a pumpkin, then it has to be a part of your theory that pumpkins are what concepts are. I take it that this is just truistic.

136

J. Fodor

Sometimes it's clear in which direction the explanation should go, and sometimes it isn't. So, for example, one's theory about having a cat ought surely to be parasitic on one's theory about being a cat; first you say what a cat is, and then you say that having a cat is just: having one of those. With jobs, pains, and siblings, however, it goes the other way round. First you say what is to have a job, or a pain, or a sibling, and then the story about what jobs, pains and siblings are is a spin-off. These examples are, I hope, untendentious. But decisions about the proper order of explanation can be unobvious, important, and extremely difficult. To cite a notorious case: ought one first to explain what the number three is and then explain what it is for a set to have three members? Or do you first explain what sets are, and then explain what numbers are in terms of them? Or are the properties of sets and of numbers both parasitic on those of something quite else (like counting, for example). If I knew and I was rich, I would be rich and famous. Anyhow, classical representational theories uniformly took it for granted that the explanation of concept possession should be parasitic on the explanation of concept individuation. First you say what it is for something to be the concept X - y o u give the "identity conditions" for the concept-and then the story about concept possession follows without further fuss. Well, but how do you identify a concept? Answer: you identify a concept by saying what it is the concept of. The concept DOG, for example, is the concept of dogs; that's to say, it's the concept that you use to think about dogs with. Correspondingly, having the concept DOG is just having a concept to think about dogs with. Similarly, mutatis mutandis, for concepts of other than canine content: the concept X is the concept of Xs. Having the concept X is just having a concept to think about Xs with. (More precisely, having the concept X is having a concept to think about Xs "as such" with. The context "thinks about . . . " is intentional for the " . . " position. We'll return to this presently.) So much for the explanatory tactics of classical representational theories of mind. Without exception, however, current theorizing about concepts reverses the classical direction of analysis. The substance of current theories lies in what they say about the conditions for having the concept X. It's the story about being the concept X - t h e story about concept individuation - that they treat as parasitic: the concept X is just whatever it is that a creature has when it has that concept. (See, for example, Peacocke, 1992, which is illuminatingly explicit on this point.) This subtle, and largely inarticulate, difference between contemporary representational theories and their classical forebears has, so I'll argue, the most profound implications for our cognitive science. To a striking extent, it determines the kinds of problems we work on and the kinds of theories that we offer as solutions to our problems. I suspect that it was a wrong turn-on balance, a catastrophe - and that we shall have to go back and do it all again.

Concepts: a potboiler

137

First, however, just a little about why the classical representational view was abandoned. There were, I think, three kinds of reasons: methodological, metaphysical and epistemological. We'll need to keep them in mind when we turn to discussing current accounts of concepts. Methodology: Suppose you're a behaviorist of the kind who thinks there are no concepts. In that case, you will feel no need for a theory about what concepts are, classical or otherwise. Behaviorist views aren't widely prevalent now, but they used to be; one of the things that killed the classical theory of concepts was simply that concepts are mental entities,3 and mentalism went out of fashion. Metaphysics: A classical theory individuates concepts by specifying their contents; the concept X is the concept of Xs. This seemed OK-it seemed not to beg any principled questions - because classical theorists thought that they had of-ness under control; they thought the image theory of mental representation explained it. We now know that they were wrong to think this. Even if concepts are mental images (which they aren't) and even if the concept DOG looks like a dog (which it doesn't) still, it isn't because it looks like a dog that it's concept of dogs. Of-ness ("content", "intentionality") does not reduce to resemblance, and it is now widely, and rightly, viewed as problematic. It doesn't follow either that classical theorists were wrong to hold that the story about concept possession should be parasitic on the story about concept identification, or that they were wrong to hold that concepts should be individuated by their contents. But it's true that if you want to defend the classical order of analysis, you need an alternative to the picture theory of meaning. Epistemology: The third of the standard objections to the classical account of concepts, though at least as influential as the others, is distinctly harder to state. Roughly, it's that classical theories aren't adequately "ecological". Used in this connection, the term has Gibsonian ring; but I'm meaning it to pick out a much broader critical tradition. (In fact, I suspect Dewey was the chief influence; see the next footnote.) Here's a rough formulation. What cognitive science is trying to understand is something that happens in the world; it's the interplay of environmental contingencies and behavioral adaptations. Viewing concepts primarily as the vehicles of thought puts the locus of this mind/world interaction (metaphorically and maybe literally) not in the world but in the head. Having put it in there, classical theorists are at a loss as to how to get it out again. So the ecological objection goes. This kind of worry comes in many variants, the epistemological being, perhaps, the most familiar. If concepts are internal mental representations, and thought is conversant only with concepts, how does thought every contact the external world Terminological footnote: here and elsewhere in this paper, I follow the psychologist's usage rather than the philosopher's; for philosophers, concepts are generally abstract entities, hence, of course, not mental. The two ways of talking are compatible. The philosopher's concepts can be viewed as the types of which the psychologist's concepts are tokens.

J. Fodor

138

that the mental representations are supposed to represent? If there is a "veil of ideas" between the mind and the world, how can the mind see the world through the veil? Isn't it, in fact, inevitable that the classical style of theorizing eventuates either in solipsism ("we never do connect with the world, only with our idea of it") or in idealism ("it's OK if we can never get outside of heads because the world is in there with us")?4 And, surely, solipsism and idealism are both refutations of theories that entail them. Notice that this ecological criticism of the classical story is different from the behaviorist's eschewal of intentionality as such. The present objection to "internal representations" is not that they are representations, but that they are internal. In fact, this sort of objection to the classical theory predates behaviorism by a lot. Reid used it against Hume, for example. Notice too that this objection survives the demise of the image theory of concepts; treating mental representation as, say, discursive rather than iconic doesn't help. What's wanted isn't either pictures of the world or stories about the world; what's wanted is what they call in Europe being in the world. (I'm told this sounds even better in German.) This is all, as I say, hard to formulate precisely; I think, in fact, that it is extremely confused. But even if the "ecological" diagnosis of what's wrong with classical concepts is a bit obscure, it's clear enough what cure was recommended, and this brings us back to our main topic. If what we want is to get thought out of the head and into the world, we need to reverse the classical direction of analysis, precisely as discussed above; we need to take having a concept as the fundamental notion and define concept individuation in terms of it. This is a true Copernican revolution in the theory of mind, and we are still living among the debris. Here, in the roughest outline, is the new theory about concept possession: having a concept is having certain epistemic capacities. To have the concept of X is to be able to recognize Xs, and/or to be able to reason about Xs in certain kinds of ways. (Compare the classical view discussed above: having the concept of X is just being able to have thoughts about Xs). It is a paradigmatically pragmatist idea that having a concept is being able to do certain things rather than being able to think certain things. Accordingly, in the discussion that follows, I will contrast classical theories of concepts with "pragmatic" ones. I'll try to make it plausible that all the recent and current accounts of concepts in cognitive science really are just variations on the pragmatist legacy. 4

"Experience to them is not only something extraneous which is occasionally superimposed upon nature, but it forms a veil or screen which shuts us off from nature, unless in some way it can be 'transcended' (p. la)". "Other [philosophers' methods] begin with results of a reflection that has already torn in two the subject-matter and the operations and states of experiencing. The problem is then to get together again what has been sundered ..." (p. 9). Thus Dewey (1958). The remedy he recommends is resolutely to refuse to recognize the distinction between experience and its object. "[Experience] recognizes in its primary integrity no division between act and material, subject and object, but contains them both in an unanalyzed totality."

Concepts: a potboiler

139

In particular, I propose to consider (briefly, you'll be pleased to hear) what I take to be five failed versions of pragmatism about concepts. Each evokes its proprietary nemesis; there is, for each, a deep fact about concepts by which it is undone. The resulting symmetry is gratifyingly Sophoclean. When we've finished with this catalogue of tragic flaws, we'll have exhausted all the versions of concept pragmatism I've heard of, or can think of, and we'll also have compiled a must-list for whatever theory of concepts pragmatism is eventually replaced by.

2.1. Behavioristic pragmatism (and the problem of intentionality) I remarked above that behaviorism can be a reason for ruling all mentalistic notions out of psychology, concepts included. However, not all behaviorists were eliminativists; some were reductionists instead. Thus Ryle, and Hull (and even Skinner about half the time) are perfectly content to talk of concept possession, so long as the "criteria" for having a concept can be expressed in the vocabulary of behavior and/or in the vocabulary of dispositions to behave. Do not ask what criteria are; there are some things we're not meant to know. Suffice it that criterial relations are supposed to be sort-of-semantical rather than sort-of-empirical. So, then, which behaviors are supposed to be criterial for concept possession? Short answer: sorting behaviors. Au fond, according to this tradition, having the concept X is being able to discriminate Xs from non-Xs; to sort things into the ones that are X and the ones that aren't. Though behaviorist in essence, this identification of possessing a concept with being able to discriminate the things it applies to survived well into the age of computer models (see, for example, "procedural" semanticists like Woods, (1975); and lots of philosophers still think there must be something to it (see, for example, Peacocke, 1992). This approach gets concepts into the world with a vengeance: having a concept is responding selectively, or being disposed to respond selectively, to the things in the world that the concept applies to; and paradigmatic responses are overt behaviors "under the control" of overt stimulations. I don't want to bore your with ancient recent history, and I do want to turn to less primitive versions of pragmatism about concepts. So let me just briefly remind you of what proved to be the decisive argument against the behavioristic version: concepts can't be just sorting capacities, for if they were, then coextensive concepts - concepts that apply to the same things-would have to be identical. And coextensive concepts aren't, in general, identical. Even necessarily coextensive concepts - like TRIANGULAR and TRILATERAL, for examplemay perfectly well be different concepts. To put this point another way, sorting is something that happens under a description; it's always relative to some or other way of conceptualizing the things that are being sorted. Though their behaviors

140

J. Fodor

may look exactly the same, and though they may end up with the very same things in their piles, the creature that is sorting triangles is in a different mental state, and is behaving in a different way, from the creature that is sorting trilaterals; and only the first is exercising the concept TRIANGLE. (For a clear statement of this objection to behaviorism, see Dennett, 1978.) Behaviorists had a bad case of mauvais fois about this; they would dearly have liked to deny the intentionality of sorting outright. In this respect, articles like Kendler (1952), according to which 'what is learned, [is] a pseudoproblem in psychology", make fascinating retrospective reading. Suppose, however, that you accept the point that sorting is always relative to a concept, but you wish, nonetheless, to cleave to some kind of pragmatist reduction of concept individuation to concept possession and of concept possession to having epistemic capacities. The question then arises: what difference in their epistemic capacities could distinguish the creature that is sorting triangles from the creature that is sorting trilaterals? What could the difference between them be, if it isn't in the piles that they end up with? The universally popular answer has been that the difference between sorting under the concept TRIANGLE and sorting under the concept TRILATERAL lies in what the sorter is disposed to infer from the sorting he performs. To think of something as a triangle is to think of it as having angles; to think of something as a trilateral is to think of it as having sides. The guy who is collecting triangles must therefore accept that the things in his collection have angles (whether or not he has noticed that they have sides); and the guy who is collecting trilaterals must accept that the things in his collection have sides (even if he hasn't notice that they have angles). The long and short is: having concepts is having a mixture of abilities to sort and abilities to infer.5 Since inferring is presumably neither a behavior nor a behavioral capacity, this formulation is, of course, not one that a behavioristic pragmatist can swallow. So much the worse for behaviorists, as usual. But notice that pragmatists as such are still OK: even if having a concept isn't just knowing how to sort things, it still may be that having a concept is some kind of knowing how, and that theories of concept possession are prior to theories of concept individuation. We are now getting very close to the current scene. All non-behaviorist 5 The idea that concepts are (at least partially) constituted by inferential capacities receives what seems to be independent support from the success of logicist treatments of the "logical" concepts (AND, ALL, etc.). For many philosophers (though not for many psychologists) thinking of concepts as inferential capacities is a natural way of extending the logicist program from the logical vocabulary to TREE or TABLE. So, when these philosophers tell you what it's like to analyze a concept, they start with AND. (Here again, Peacocke, 1992, is paradigmatic.) It should, however, strike you as not obvious that the analysis of AND is a plausible model for the analysis of TREE or TABLE.

Concepts: a potboiler

141

versions of pragmatism hold that concept possession is constituted, at least in part, by inferential dispositions and capacities. They are thus all required to decide which inferences constitute which concepts. Contemporary theories of concepts, though without exception pragmatist, are distinguished by the ways that they approach this question. Of non-behavioristic pragmatist theories of concepts there are, by my reckoning, exactly four. Of which the first is as follows.

2.2. Anarchic pragmatism (and the realism problem) Anarchic pragmatism is the doctrine that though concepts are constituted by inferential dispositions and capacities, there is no fact of the matter about which inferences constitute which concepts. California is, of course, the locus classicus of anarchic pragmatism; but no doubt there are those even on the East Coast who believe it in their hearts. I'm not going to discuss the anarchist view. If there are no facts about which inferences constitute which concepts, then there are no facts about which concepts are which. And if there are no facts about which concepts are which, then there are no facts about which beliefs and desires are which (for, by assumption, beliefs and desires are complexes of which concepts are the constituents). And if there are no facts about which beliefs and desires are which, there is no intentional cognitive science, for cognitive science is just belief/desire explanation made systematic. And if there is no cognitive science, we might as well stop worrying about what concepts are and have a nice long soak in a nice hot tub. I'm also not going to consider a doctrine that is closely related to anarchic pragmatism: namely, that while nothing systematic can be said about concept identity, it may be possible to provide a precise account of when, and to what degree, two concepts are similar. Some such thought is often voiced informally in the cognitive science literature, but there is, to my knowledge, not even a rough account of how such a similarity relation over concepts might be defined. I strongly suspect this is because a robust notion of similarity is possible only where there is a correspondingly robust notion of identity. For a discussion, see Fodor and Lepore (1992, Ch. 7).

2.3. Definitional pragmatism (and the analyticity problem) Suppose the English word "bachelor" means the same as the English phrase "unmarried male". Synonymous terms presumably express the same concept (this is a main connection between theories about concepts and theories about language), so it follows that you couldn't have the concept BACHELOR and fail to have the concept UNMARRIED MALE. And from that, together with the

142

J. Fodor

intentionality of sorting (see section 2.1), it follows that you couldn't be collecting bachelors so described unless you take yourself to be collecting unmarried males; that is, unless you accept the inference that if something belongs in your bachelor collection, then it is something that is male and unmarried. Maybe this treatment generalizes; maybe, having the concept X just is being able to sort Xs and being disposed to draw the inferences that define X-ness. The idea that it's defining inferences that count for concept possession is now almost as unfashionable as behaviorism. Still, the departed deserves a word or two of praise. The definition story offered a plausible (though partial) account of the acquisition of concepts. If BACHELOR is the concept UNMARRIED MALE, then it's not hard to imagine how a creature that has the concept UNMARRIED and has the concept MALE could put them together and thereby achieve the concept BACHELOR. (Of course the theory that complex concepts are acquired by constructing them from their elements presupposes the availability of the elements. About the acquisition of these, definitional pragmatism tended to be hazy.) This process of assembling concepts can be-indeed, was-studied in the laboratory; see Bruner, Goodnow, & Austin (1956) and the large experimental literature that it inspired. Other significant virtues of the definition story will suggest themselves when we discuss concepts as prototypes in section 2.4. But alas, despite its advantages, the definition theory doesn't work. Concepts can't be definitions because most concepts don't have definitions. At a minimum, to define a concept is to provide necessary and sufficient conditions for something to be in its extension (i.e., for being among the things that concept applies to). And, if the definition is to be informative, the vocabulary in which it is couched must not include either the concept itself or any of its synonyms. As it turns out, for most concepts, this condition simply can't be met; more precisely, it can't be met unless the definition employs synonyms and near-synonyms of the concept to be defined. Maybe being male and unmarried is necessary and sufficient for being a bachelor; but try actually filling in the blanks in "JC is a dog iff JC is a ..." without using the words like "dog" or "canine" or the like on the right-hand side. There is, to be sure, a way to do it; if you could make a list of all and only the dogs (Rover, Lassie, Spot . . . etc.), then being on the list would be necessary and sufficient for being in the extension of DOG. That there is this option is, however, no comfort for the theory that concepts are definitions. Rather, what it shows is that being a necessary and sufficient condition for the application of a concept is not a sufficient condition for being a definition of the concept. This point generalizes beyond the case of lists. Being a creature with a backbone is necessary and sufficient for being a creature with a heart (so they tell me). But it isn't the case that "creature with a backbone" defines "creature with a heart" or vice versa. Quite generally, it seems that Y doesn't define X unless Y applies to all and only the possible Xs (as well, of course, as all and only the

Concepts: a potboiler

143

actual Xs). It is, then, the modal notion - possibility - that's at the heart of the idea that concepts are definitions. Correspondingly, what killed the definition theory of concepts, first in philosophy and then in cognitive psychology, is that nobody was able to explicate the relevant sense of "possible". It seems clear enough that even if Rover, Lassie and Spot are all the dogs that there actually are, it is possible, compatible with the concept of DOG, that there should be others; that's why you can't define DOG by just listing the dogs. But is it, in the same sense, possible, compatible with the concept DOG that some of these non-actual dogs are ten feet long? How about twenty feet long? How about twenty miles long? How about a light-year long? To be sure, it's not biologically possible that there should be a dog as big as a light-year; but presumably biology rules out a lot of options that the concept DOG, as such, allows. Probably biology rules out zebra-striped dogs; surely it rules out dogs that are striped red, white and blue. But I suppose that red, white and blue striped dogs are conceptually possible; somebody who thought that there might be such dogs wouldn't thereby show himself not to have the concept DOG - would he? So, again, are light-year-long dogs possible, compatible with the concept DOG? Suppose somebody thought that maybe there could be a dachshund a light-year long. Would that show that he has failed to master the concept DOG? Or the concept LIGHT-YEAR? Or both? To put the point in the standard philosophical jargon: even if light-year-long dogs aren't really possible, "shorter than a light-year" is part of the definition of DOG only if "some dogs are longer than a light-year" is analytically impossible; mere biological or physical (or even metaphysical) impossibility won't do. Well, is it analytically impossible that there should be such dogs? If you doubt that this kind of question has an answer, or that it matters a lot for any serious purpose what the answer is, you are thereby doubting that the notion of definition has an important role to play in the theory of concept possession. So much for definitions.

2.4. Stereotypes and prototypes (and the problem of compositionality) Because it was pragmatist, the definition story treated having a concept as having a bundle of inferential capacities, and was faced with the usual problem about which inferences belong to which bundles. The notion of an analytic inference was supposed to bear the burden of answering this question, and the project foundered because nobody knows what makes an inference analytic, and nobody has any idea how to find out. "Well", an exasperated pragmatist might nonetheless reply, "even if I don't know what makes an inference analytic, I do know what makes an inference statistically reliable. So why couldn't the theory of concept possession be statistical rather than definitional? Why couldn't I exploit

144

J. Fodor

the notion of a reliable inference to do what definitional pragmatism tried and failed to do with the notion of an analytic inference?" We arrive, at last, at modern times. For lots of kinds of Xs, people are in striking agreement about what properties an arbitrarily chosen X is likely to have. (An arbitrarily chosen bird is likely to be able to fly; an arbitrarily chosen conservative is likely to be a Republican; an arbitrarily chosen dog is likely to be less than a light-year long.) Moreover, for lots of kinds of Xs, people are in striking agreement about which Xs are prototypic of the kind (diamonds for jewels; red for colors; not dachshunds for dogs). And, sure enough, the Xs that are judged to be prototypical are generally ones that have lots of the properties that an arbitrary X is judged likely to have; and the Xs that are judged to have lots of the properties that an arbitrary X is likely to have are generally the ones that are judged to be prototypical. Notice, in passing, that stereotypes share one of the most agreeable features of definitions: they make the learning of (complex) concepts intelligible. If the concept of an X is the concept of something that is reliably Y and Z, then you can learn the concept X if you have the concepts Y and Z together with enough statistics to recognize reliability when you see it. It would be OK, for this purpose, if the available statistical procedures were analogically (rather than explicitly) represented in the learner. Qua learning models, "neural networks" are analog computers of statistical dependencies, so it's hardly surprising that prototype theories of concepts are popular among connectionists. (See, for example, McClelland & Rummelhart, 1986.) So, then, why shouldn't having the concept of an X be having the ability to sort by X-ness, together with a disposition to infer from something's being X to its having the typical properties of Xs? I think, in fact, that this is probably the view of concepts that the prototypical cognitive scientist holds these days. To see why it doesn't work, let's return one last time to the defunct idea that concepts are definitions. It was a virtue of that idea that it provides for the compositionality of concepts, and hence for the productivity and systematicity of thought. This, we're about to see, is no small matter. In the first instance, productivity and systematicity are best illustrated by reference to features (not of minds but) of natural languages. To say that languages are productive is to say that there is no upper bound to the number of well-formed formulas that they contain. To say that they are systematic is to say that if a language can express the proposition that P, then it will be able to express a variety of other propositions that are, in one way or another, semantically related to P. (So, if a language can say that P and that - Q , it will also be able to say that Q and that - P ; if it can say that John loves Mary, it will be able to say that Mary loves John . . . and so forth.) As far as anybody knows, productivity and systematicity are universal features of human languages. Productivity and systematicity are also universal features of human thought

Concepts: a potboiler

145

(and, for all I know, of the thoughts of many infra-human creatures). There is no upper bound to the number of thoughts that a person can think. (I am assuming the usual distinctions between cognitive "competence" and cognitive "performance"). And also, if a mind can entertain the thought that P and any negative thoughts, it can also entertain the thought that - P ; if it can entertain the thought that Mary loves John, it can entertain the thought that John loves Mary . . . and so on. It is extremely plausible that the productivity and the systematicity of language and thought are both to be explained by appeal to the systematicity and productivity of mental representations, and that mental representations are systematic and productive because they are compositional. The idea is that mental representations are constructed by the application of a finite number of combinatorial principles to a finite basis of (relatively or absolutely) primitive concepts. (So, the very same process that gets you from the concept MISSILE to the concept ANTIMISSILE, also gets you from the concept ANTIMISSILE to the concept ANTIANTIMISSLE, and so on ad infinitum.) Productivity follows because the application of these constructive principles can iterate without bound. Systematicity follows because the concepts and principles you need to construct the thoughts that P and -Q are the very same ones that you need to construct the thoughts that Q and - P ; and the concepts and principles you need to construct the thought that John loves Mary are the very same ones that you need to construct the thought that Mary loves John. This sort of treatment of compositionality is familiar, and I will assume that it is essentially correct. I want to emphasize that it places a heavy constraint on both theories of concept possession and theories of concept individuation. If you accept compositionality, then you are required to say that whatever the concept DOG is that occurs in the thought that Rover is a dog, that very same concept DOG also occurs in the thought that Rover is a brown dog; and, whatever the concept BROWN is that occurs in the thought that Rover is brown, the very same concept BROWN also occurs in the thought that Rover is a brown dog. It's on these assumptions that compositionality explains how being able to think that Rover is brown and that Rover is a dog is linked to being able to think that Rover is a brown dog. Compositionality requires, in effect, that constituent concepts must be insensitive to their host; a constituent concept contributes the same content to all the complex representations it occurs in. And compositionality further requires that the content of a complex representation is exhausted by the contributions that its constituents make. Whatever the content of the concept of BROWN DOG may be, it must be completely determined by the content of the constituent concepts BROWN and DOG, together with the combinatorial apparatus that sticks these constituents together; if this were not the case, your grasp of the concepts BROWN and DOG wouldn't explain your grasp of the concept BROWN DOG.

146

J. Fodor

In short, when complex concepts are compositional, the whole must not be more than the sum of its parts, otherwise compositionality won't explain productivity and systematicity. And if compositionality doesn't, nothing will. If this account of compositionality strikes you as a bit austere, it may be some comfort that the systematicity and productivity of thought is compatible with compositionality failing in any finite number of cases. It allows, for example, that finitely many thoughts (hence a fortiori, finitely many linguistic expressions) are idiomatic or metaphoric, so long as there are infinitely many that are neither. We can now see why, though concepts might have turned out to be definitions, they couldn't possibly turn out to be stereotypes or prototypes. Concepts do contribute their defining properties to the complexes of which they are constituents, and the defining properties of complex concepts are exhaustively determined by the defining properties that the constituents contribute. Since bachelors are, by definition, unmarried men, tall bachelors are, by the same definition, tall unmarried men; and very tall bachelors are very tall unmarried men, and very tall bachelors from Hoboken are very tall unmarried men from Hoboken . . . and so on. Correspondingly, there is nothing more to the definition of "very tall bachelor from Hoboken" than very tall unmarried man from Hoboken; that is, there is nothing more to the definition of the phrase than what the definitions of its constituents contribute. So, then, if concepts were definitions, we could see how thought could be compositional, and hence productive and systematic. Concepts aren't definitions, of course. It's just that, from the present perspective, it's rather a pity that they're not. For stereotypes, alas, don't work the way that definitions do. Stereotypes aren't compositional. Thus, "ADJECTIVE X" can be a perfectly good concept even if there is no adjective X stereotype. And even if there are stereotypic adjective Xs, they don't have to be stereotypic adjectives or stereotypic Xs. I doubt, for example, that there is a stereotype of very tall men from Hoboken; but, even if there were, there is no reason to suppose that it would be either a stereotype for tall men, or a stereotype for men from Hoboken, or a stereotype for men. On the contrary: often enough, the adjective in "ADJECTIVE X" is there precisely to mark a way that adjective X$ depart from stereotypic Xs. Fitzgerald made this point about stereotypes to Hemingway when he said, "The rich are different from the rest of us." Hemingway replied by making the corresponding point about definitions: "Yes", he said, "they have more money". In fact, this observation about the uncompositionality of stereotypes generalizes in a way that seems to me badly to undermine the whole pragmatist program of identifying concept possession with inferential dispositions. I've claimed that knowing what is typical of adjective and what is typical of X doesn't, in the general case, tell you what is typical of adjective Xs. The reason it doesn't is perfectly clear; though some of your beliefs about adjective Xs are compositional-

Concepts: a potboiler

147

ly inherited from your beliefs about adjectives, and some are compositionally inherited from your beliefs about Xs, some are beliefs that you have acquired about adjective Xs as such, and these aren't compositional at all. The same applies, of course, to the inferences that your beliefs about adjective Xs dispose you to draw. Some of the inferences you are prepared to make about green apples follow just from their being green and from their being apples. That is to say: they derive entirely from the constituency and structure of your GREEN APPLE concept. But others depend on information (or misinformation) that you have picked up about green apples as such: that green apples go well in apple pie; that they are likely to taste sour; that there are kinds of green apples that you'd best not eat uncooked, and so forth. Patently, these inferences are not definitional and not compositional; they are not ones that GREEN APPLE inherits from its constituents. They belong to what you know about green applies, not to what you know about the corresponding words or concepts. You learned that "green apple" means green and apple when you learned English at your mother's knee. But probably you learned that green apples mean apple pies from the likes of Julia Child. The moral is this: the content of complex concepts has to be compositionally determined, so whatever about concepts is not compositionally determined is therefore not their content. But, as we've just been seeing, the inferential role of a concept is not, in general, determined by its structure together with the inferential roles of its constituents. That is, the inferential roles of concepts are not, in general, compositional; only defining inferences are. This puts your paradigmatic cognitive scientist in something of a pickle. On the one hand, he has (rightly, I think) rejected the idea that concepts are definitions. On the other hand, he cleaves (wrongly, I think) to the idea that having concepts is having certain inferential dispositions. But, on the third hand (as it were), only defining inferences are compositional so if there are no definitions, then having concepts can't be having inferential capacities. I think that is very close to being a proof that the pragmatist notion of what it is to have a concept must be false. This line of argument was first set out in Fodor and Lepore (1992). Philosophical reaction has been mostly that if the price of the pragmatist account of concepts is reviving the notion that there are analytic/ definitional inferences, then there must indeed be analytic/definitional inferences. My own view is that cognitive science is right about concepts not being definitions, and that it's the analysis of having concepts in terms of drawing inferences that is mistaken. Either way, it seems clear that the current situation is unstable. Something's gotta give. I return briefly to my enumeration of the varieties of pragmatist theories of concept possession. It should now seem unsurprising that none of them work. In light of the issues about compositionality that we've just discussed, it appears there are principled reasons why none of them could.

148

J. Fodor

2.5. The "theory theory" of concepts (and the problem of holism) Pragmatists think that having a concept is having certain epistemic capacities; centrally it's having the capacity to draw certain inferences. We've had trouble figuring out which inferences constitute which concepts; well, maybe that's because we haven't been taken the epistemic bit sufficiently seriously. Concepts are typically parts of beliefs; but they are also, in a different sense of "part", typically parts of theories. This is clearly true of sophisticated concepts like ELECTRON, but perhaps it's always true. Even every-day concepts like HAND or TREE or TOOTHBRUSH figure in complex, largely inarticulate knowledge structures. To know about hands is to know, inter alia, about arms and fingers; to know about toothbrushes is, inter alia, to know about teeth and the brushing of them. Perhaps, then, concepts are just abstractions from such formal and informal knowledge structures. On this view, to have the concept ELECTRON is to know what physics has to say about electrons; and to have the concept TOOTHBRUSH is to know what dental folklore has to say about teeth. Here are some passages in which the developmental cognitive psychologist Susan Carey (1985) discusses the approach to concepts that she favors: "... [young] children represent only a few theory-like cognitive structures, in which their notions of causality are embedded and in terms of which their deep ontological commitments are explicated. Cognitive development consists, in part, in the emergence of new theories out of these older ones, with the concomitant reconstructing of the ontologically important concepts and emergence of new explanatory notions" (p. 14); "... successive theories differ in three related ways: in the domain of phenomena accounted for, the nature of explanations deemed acceptable, and even in the individual concepts at the center of each system . Change of one kind cannot be understood without reference to the changes of the other kinds" (pp. 4-5). The last two sentences are quoted from Carey's discussion of theory shifts in the history of science; her proposal is, in effect, that these are paradigms for conceptual changes in ontogeny. Cognitive science is where philosophy goes when it dies. The version of pragmatism according to which concepts are abstractions from knowledge structures corresponds exactly to the version of positivism according to which terms like "electron" are defined implicitly by reference to the theories they occur in. Both fail, and for the same reasons. Suppose you have a theory about electrons (viz. that they are X) and I have a different theory about electrons (viz. that they are Y). And suppose, in both cases, that our use of the term "electron" is implicitly defined by the theories we espouse. Well, the "theory theory" says that you have an essentially different concept of electrons from mine if (and only if?) you have an essentially different theory of electrons from mine. The problem of how to individuate concepts thus reduces to the problem of how to individuate theories, according to this view.

Concepts: a potboiler

149

But, of course, nobody knows how to individuate theories. Roughly speaking, theories are bundles of inferences, just as concepts are according to the pragmatist treatment. The problem about which inferences constitute which concepts has therefore an exact analagon in the problem which inferences constitute which theories. Unsurprisingly, these problems are equally intractable. Indeed, according to the pragmatist view, they are interdefined. Theories are essentially different if they exploit essentially different concepts; concepts are essentially different if they are exploited by essentially different theories. It's hard to believe it matters much which of these shells you keep the pea under. One thing does seem clear: if your way out of the shell game is to say that a concept is constituted by the whole of the theory it belongs to, you will pay the price of extravagant paradox. For example: it turns out that you and I can't disagree about dogs, or electrons, or toothbrushes since we have no common conceptual apparatus in which to couch the disagreement. You utter "Some dogs have tails." "No dogs have tails" I reply. We seem to be contradicting one another, but in fact we're not. Since tailessness is part of my theory of dogs, it is also part of my concept DOG according to the present, holist account of concept individuation. Since you and I have different concepts of dogs, we mean different things when we say "dog". So the disagreement between us is, as comfortable muddleheads like to put it, "just semantic". You might have thought that our disagreement was about the facts and that you could refute what I said by producing a dog with a tail. But it wasn't and you can't, so don't bother trying; you have you idea of dogs and I have mine. (What, one wonders, makes them both ideas of dogs?) First the pragmatist theory of concepts, then the theory theory of concepts, then holism, then relativism. So it goes. Or so, at least, it's often gone. I want to emphasize two caveats. The first is that I'm not accusing Carey of concept holism, still less of the slide from concept holism to relativism. Carey thinks that only the "central" principles of a theory individuate its concepts. The trouble is that she has no account of centrality, and the question "which of the inferences a theory licenses are central?" sounds suspiciously similar to the question "which of the inferences that a concept licenses are constitutive?" Carey cites with approval Kuhn's famous distinction between theory changes that amount to paradigm shifts and those that don't (Kuhn, 1962). If you have caught onto how this game is played, you won't be surprised to hear that nobody knows how to individuate paradigms either. Where is this buck going to stop? My second caveat is that holism about the acquisition of beliefs and about the confirmation of theories might well both be true even if holism about the individuation of concepts is, as I believe, hopeless. There is no contradiction between Quine's famous dictum that it's only as a totality that our beliefs "face the tribunal of experience", and Hume's refusal to construe the content of one's concepts as being determined by the character of one's theoretical commitments.

150

J. Fodor

There is, to be sure, a deep, deep problem about how to get a theory of confirmation and belief fixation if you are an atomist about concepts. But there is also a deep, deep problem about how to get a theory of confirmation and belief fixation if you are not an atomist about concepts. So far as I know, there's no reason to suppose that the first of these problems is worse than the second. So much for caveats. It's worth noticing that the holistic account of concepts at which we've now dead-ended is diametrically opposite to the classical view that we started with. We saw that, for the likes of Hume, any concept could become associated to any other. This was a way of saying that the identity of a concept is independent of the theories one holds about the things that fall under it; it's independent, to put it contemporary terms, of the concept's inferential role. In classical accounts, concepts are individuated by what they are concepts of, and not by what theories they belong to. Hume was thus a radical atomist just where contemporary cognitive scientists are tempted to be radically holist. In this respect, I think that Hume was closer to the truth than we are. Here's how the discussion has gone so far: modern representational theories of mind are devoted to the pragmatist idea that having concepts is having epistemic capacities. But not just sorting capacities since sorting is itself relativized to concepts. Maybe, then, inferential capacities as well? So be it, but which inferential capacities? Well, at a minimum, inferential capacities that respect the compositionality of mental representations. Defining inferences are candidates since they do respect the compositionality of mental representations. Or, rather, they would if there were any definitions, but there aren't any definitions to speak of. Statistical inferences aren't candidates because they aren't compositional. It follows that concepts can't be stereotypes. The "theory theory" merely begs the problem it is meant to solve since the individuation of theories presupposes the individuation of the concepts they contain. Holism would be a godsend and the perfect way out except that it's preposterous on the face of it. What's left, then, for a pragmatist to turn to? I suspect, in fact, that there is nothing left for a pragmatist to turn to and that our cognitive science is in deep trouble. Not that there aren't mental representations, or that mental representations aren't made of concepts. The problem is, rather, that Hume was right: concepts aren't individuated by the roles that they play in inferences, or, indeed, by their roles in any other mental processes. If, by stipulation, semantics is about what constitutes concepts and psychology is about the nature of mental processes, then the view I'm recommending is that semantics isn't part of psychology. If semantics isn't part of psychology, you don't need to have a sophisticated theory of mental processes in order to get it right about what concepts are. Hume, for example, did get it right about what concepts are, even though his theory of mental processes was associationistic and hence hopelessly primitive. Concepts are the constituents of thoughts; as such, they're the most elementary mental

Concepts: a potboiler

151

objects that have both causal and representational properties. Since, however, concepts are individuated by their representational and not by their casual properties, all that has to specified in order to identify a concept is what it is the concept of. The whole story about the individuation of the concept DOG is that it's the concept that represents dogs, as previously remarked. But if "What individuates concepts?" is easy, that's because its the wrong question, according to the present view. The right questions are: "How do mental representations represent?" and "How are we to reconcile atomism about the individuation of concepts with the holism of such key cognitive processes as inductive inference and the fixation of belief?" Pretty much all we know about the first question is that here Hume was, for once, wrong; mental representation doesn't reduce to mental imaging. What we know about the second question is, as far as I can tell, pretty nearly nothing at all. The project of constructing a representational theory of the mind is among the most interesting that empirical science has ever proposed. But I'm afraid we've gone about it all wrong. At the very end of Portnoy's Complaint, the client's two hundred pages of tortured, non-directive self-analysis comes to an end. In the last sentence of the book, the psychiatrist finally speaks: "So [said the doctor]. Now vee may perhaps to begin. Yes?"

References Bruner, J., Goodnow, J., & Austin, G. (1956). A Study of Thinking. New York: Wiley. Carey, S. (1985). Conceptual Change in Childhood. Cambridge, MA: MIT Press. Dennett, D. (1978). Skinner Skinned. In Brainstorms. Cambridge, MA: MIT Press. Dewey, J. (1958). Experience and Nature. New York: Dover Publications. Fodor, J. (1990). A Theory of Content and Other Essays. Cambridge, MA: MIT Press. Fodor, J., & Lepore, E. (1991). Why meaning (probably) isn't conceptual role. Mind and Language, 6, 328-343. Fodor, J., & Lepore, E. (1992). Holism: A Shopper's Guide. Oxford: Blackwell. Fodor, J., & McLaughlin, B. (1990). Connectionism and the problem of systematicity: why Smolensky's solution doesn't work. Cognition, 35, 183-204. Fodor, J., & Pylyshyn, Z. (1981). How direct is visual perception? Some reflection on Gibson's "ecological approach". Cognition, 9, 139-196. Fodor, J., & Pylyshyn, Z. (1988). Correctionism and cognitive architecture. Cognition, 28, 3-71. Kendler, H. (1952). "What is learned?" A theoretical blind alley. Psychological Review, 59, 269-277. Kuhn, T. (1962). The Structure of Scientific Revolutions. Chicago: University of Chicago Press. McClelland, J., & Rummelhart, D. (1986). A distributed model of human learning and memory. In J. McClelland & D. Rummelhart (Eds.), Parallel Distributed Processing (Vol. 2). Cambridge, MA: MIT Press. Peacocke, C. (1992). A Study of Concepts. Cambridge, MA: MIT Press. Smolensky, P. (1988). On the proper treatment of connectionism. Behavioral and Brain Sciences, 11, 1-23. Woods, W. (1975). What's in a link? In D. Bobrow & A. Collins (Eds.), Representation and Understanding. New York: Academic Press.

8

Young children's naive theory of biology Giyoo Hatano*' a , Kayoko Inagaki b ^Faculty of Liberal Arts, Dokkyo University, Soka, Saitama 340, Japan b School of Education, Chiba University, Chiba 263, Japan

Abstract This article aimed at investigating the nature of young children's naive theory of biology by reviewing a large number of studies conducted in our laboratories. More specifically, we tried to answer the following five critical questions. What components does young children's knowledge system for biological phenomena (or naive biology) have? What functions does it have in children's lives? How is it acquired in ontogenesis? How does its early version change as children grow older? Is it universal across cultures and through history? We propose that young children's biological knowledge system has at least three components, that is, knowledge needed to specify the target objects of biology, ways of inferring attributes or behaviors of biological kinds, and a non-intentional causal explanatory framework, and that these three constitute a form of biology, which is adaptive in children's lives. We also claim that the core of naive biology is acquired based on specific cognitive constraints as well as the general mechanism of personification and the resultant vitalistic causality, but it is differently instantiated and elaborated through activity-based experiences in the surrounding culture.

Introduction A growing number of cognitive developmentalists have come to agree that young children possess "theories" about selected aspects of the world (Wellman & Gelman, 1992). This conceptualization is a distinct departure from the Piagetian position, which assumed young children to be preoperational and thus

* Corresponding author.

154

G. Hatano, K. Inagaki

incapable of offering more or less plausible explanations in any domain, because the term "theories" means coherent bodies of knowledge that involve causal explanatory understanding. How similar young children's theories are to theories scientists have is still an open question, but the former certainly have something more than a collection of facts and/or procedures to obtain desired results (Kuhn, 1989). An important qualification here is "selected aspects of". In other words, young children are assumed to possess theories only in a few selected domains, where innate or early cognitive constraints work. Carey (1987) suggest that there are a dozen or so such domains. It is generally agreed that naive physics and naive psychology are included among them. What else? Wellman and Gelman (1992) take biology as the third domain. As to whether young children have acquired a form of biology, however, there has been a debate in recent years. On one hand, Carey (1985) claimed that children before around age 10 make predictions and explanations for biological phenomena based on intuitive psychology (i.e., intentional causality). According to her, young children lack the mind-body distinction, more specifically, do not recognize that our bodily functions are independent of our intention nor that biological processes which produce growth or death are autonomous. On the other hand, a number of recent studies have suggested that children possess biological knowledge at much earlier ages than Carey claimed. Some developmentalists (e.g., Hatano & Inagaki, 1987) have asserted that the differentiation between psychology and biology occurs, if it does, much earlier than Carey (1985) assumed. Others have proposed that biological phenomena are conceptualized differently from other phenomena from the beginning (e.g., Keil, 1992). A few other candidate theories young children may possess are theory of matters (e.g., Smith, Carey, & Wiser, 1985), astronomy (e.g., Vosniadou & Brewer, 1992), and theory of society (e.g., Furth, 1980). Nonetheless, none of these has been widely accepted as an important domain, nor researched extensively, at least compared with the "big three". Whichever aspects of the world young children have theories about, exact characterizations of these theories require further studies. Among others, the following questions seem critical. What components does each theory have? What functions does it have in children's lives? How is it acquired in ontogenesis? How does its early version change as children grow older? Is it universal across cultures and through history? In what follows, we would like to offer our tentative answers to these questions as to naive biology, based on a large amount of data collected by our associates and ourselves. Because of the limited space available, we will refer to studies conducted in other laboratories only when they are highly relevant.

Young children's naive theory of biology

755

Components of young children's biological knowledge We are convinced that the body of knowledge which young children possess about biological phenomena (e.g., behavior of animals and plants needed for individual survival; bodily process; reproduction and inheritance of properties to offspring) has at least three components, and believe that these three constitute a naive biology (Inagaki, 1993b). The first is knowledge enabling one to specify objects to which biology is applicable; in other words, knowledge about the living-non-living distinction, and also about the mind-body distinction. The second is a mode of inference which can produce consistent and reasonable predictions for attributes or behaviors of biological kinds. The third is a nonintentional causal explanatory framework for behaviors needed for individual survival and bodily processes. These components correspond respectively to the three features that Wellman (1990) lists in characterizing framework theories: ontological distinctions, coherence, and a causal-explanatory framework.

Animate-inanimate and mind-body distinctions An increasing number of recent studies have revealed that young children have the animate-inanimate distinction. More specifically, preschool children can distinguish animals from inanimate objects by attending to some salient distinctive features, for example animals' ability to perform autonomous movements (e.g., Gelman, 1990). Though only a small number of studies have dealt with plants as living things, they have also indicated that young children recognize plants as distinct from non-living things in some respects. For example, children before age 6 distinguish plants and animals from non-living things in terms of growth, that is, changes in size as time goes by (Inagaki, 1993a). In this study, which is an extension of that of Rosengren, Gelman, Kalish, and McCormick (1991), which investigated children's differentiation between animals and artifacts in terms of growth, children of ages 4-6 were presented with a picture of a flower's bud (or a new artifact or a young animal) as the standard stimulus picture, and were then asked to choose which of two other pictures would represent the same plant (or artifact or animal) a few hours later and several months/years later. The children showed "invariance" patterns (i.e., no change in size both a few hours later and several months/years later for all the items) for artifacts, but "growth" patterns (i.e., changes in size either/both a few hours later and several months/years later) for plants and animals. Backscheider, Shatz, and Gelman (1993) also reported that 4-year-olds recognize that, when damaged, both animals and plants can regrow, whereas artifacts can be mended only by human intervention. It is clear

156

G. Hatano, K. Inagaki

that young children can distinguish typical animals and plants from typical non-living things in some attributes. That young children treat inanimate things differently from animals and plants is not sufficient for claiming that they have an integrated category of living things. Proof that they are aware of the commonalities between animals and plants is needed. By asking 5- and 6-year-olds whether a few examples of plants or those of inanimate things would show similar phenomena to those we observe for animals, Hatano and Inagaki (1994) found that a great majority of them recognized commonalities between animals and plants in terms of feeding and growing in size over time, and thus distinguished them from inanimate things. Moreover, many of them justified their responses by mapping food for animals to water for plants, such as "A tulip or a pine tree dies if we do not water it"; for growth/about one-third of them cited the phenomenon of plants getting bigger from a seed or a bud, and one-fifth of them by referring to watering as corresponding to feeding as a condition for growth. Based on this and other related studies, we can conclude that children are able to acquire the living-non-living distinction by age 6. Young children can also distinguish between the body and the mind, in other words, biological phenomena from social or psychological ones, both of which are observed among a subset of animate things. Springer and Keil (1989) reported that children of ages 4-7 consider those features leading to biologically functional consequences for animals to be inherited, while other sorts of features, such as those leading to social or psychological consequences, to be not. Siegal (1988) indicated that children of ages 4-8 recognize that illness is caused not by moral but by medical factors; they have substantial knowledge of contagion and contamination as causes of illness. Inagaki and Hatano (1987) revealed that children of 5-6 years of age recognize that the growth of living things is beyond their intentional control. For example, a baby rabbit grows not because its owner wants it to but because it takes food. These findings all suggest that young children recognize the autonomous nature of biological processes. An even more systematic study on the mind-body distinction has just been reported by Inagaki and Hatano (1993, Experiment 1). By interviewing children using a variety of questions, they showed that even children aged 4 and 5 already recognize not only the differential modifiability among characteristics that are unmodifiable by any means (e.g., gender), that are bodily and modifiable by exercise or diet (e.g., running speed), and that are mental and modifiable by will or monitoring (e.g., forgetfulness), but also the independence of activities of bodily organs (e.g., heartbeat) from a person's intention. Another important piece of evidence for this distinction is young children's use of non-intentional (or vitalistic) causality for bodily phenomena but not for social-psychological ones; this point is discussed in a later section.

Young children's naive theory of biology

157

Personification as means to make educated guesses about living things When children do not have enough knowledge about a target animate object, they can make an educated guess by using personification or the person analogy in a constrained way. Young children are so familiar with humans that they can use their knowledge about humans as a source for analogically attributing properties to less familiar animate objects or predicting the reactions of such objects to novel situations, but they do not use knowledge about humans indiscriminately. Our studies (Inagaki & Hatano, 1987, 1991) confirmed such a process of constrained personification. In one of the studies (Inagaki & Hatano, 1991), we asked children of age 6 to predict a grasshopper's or tulip's reactions to three types of novel situations: (a) similar situations, in which a human being and the target object would behave similarly; (b) contradictory situations, where the target object and a human would react differently, and predictions based on the person analogy contradict children's specific knowledge about the target; and (c) compatible situations, where the object and a human being would in fact react differently, but predictions obtained through the person analogy do not seem implausible to them. Example questions for each situation are as follows: "We usually feed a grasshopper once or twice a day when we raise it at home. What will happen with it if we feed it 10 times a day?" (similar situation); "Suppose a woman buys a grasshopper. On her way home she drops in at a store with this caged grasshopper. After shopping she is about to leave the store without the grasshopper. What will the grasshopper do?" (contradictory); "Does a grasshopper feel something if the person who has been taking care of it daily dies? [If the subject's answer is "Yes"] How does it feel?" (compatible). Results indicated that for the similar situations many of the children generated reasonable predictions with some explanations by using person analogies, whereas they did not give personified predictions for the contradictory situations. As expected, they produced unreasonable predictions for the compatible situations, where they were unable to check the plausibility of products of person analogies because of the lack of adequate knowledge (e.g., about the relation between the brain and feeling). The following are example responses of a child aged 6 years 3 months: for the "too-much-eating" question of the "similar" situation, "The grasshopper will be dizzy and die, 'cause the grasshopper, though it is an insect, is like a person (in this point)"; for the "left-behind" question of the "contradictory" situation, "The grasshopper will be picked up by someone, 'cause it cannot open the cage." ["// someone does not pick up the cage, what will the grasshopper doT] "The grasshopper,will just stay there." ["Why doesn't the grasshopper do anything? Why does it just stay thereT] "It cannot (go out of the cage and) walk, unlike a

158

G. Hatano, K. Inagaki

person"; for the "caretaker's death" question of the "compatible" situation, "The grasshopper will feel unhappy." This illustrates well how this child applied knowledge about humans differentially according to the types of situations. Generally speaking, children generate reasonable predictions, using person analogies in a constrained way, and the person analogy may be misleading only where they lack biological knowledge to check analogy-based predictions.

Non-intentional causality The experimental evidence presented so far enables us to indicate that young children have a coherently organized body of knowledge applicable to living things. This body of knowledge can be called a theory only when a causal explanatory framework is included in it. This concerns the third component of their biological knowledge. Here the type of causality, intentional or nonintentional, determines the nature of a theory. Carey (1985) claimed that, as mentioned above, children before age 10 base their explanations of biological phenomena on an intentional causality, because they are ignorant of physiological mechanisms involved. On the contrary, we claim that young children before schooling can apply a non-intentional causality in explaining biological phenomena, and thus they have a form of biology which is differentiated from psychology. Young children cannot give articulated mechanical explanations when asked to explain biological phenomena (e.g., bodily processes mediating input-output relations) in an open-ended interview (e.g., Gellert, 1962); sometimes they try to explain them using the language of person-intentional causality (Carey, 1985). These findings apparently support the claim that young children do not yet have biology as an autonomous domain. It seems inevitable to accept this claim so long as we assume only two types of causalities, that is, intentional causality versus mechanical causality, as represented by Carey (1985). However, we propose an intermediate form of causality between these two. Children may not be willing to use intentional causality for biological phenomena but not as yet able to use mechanical causality. These children may rely on this intermediate form of causality, which might be called "vitalistic causality". Intentional causality means that a person's intention causes the target phenomenon, whereas mechanical causality means that physiological mechanisms cause the target phenomenon. For instance, a specific bodily system enables a person, irrespective of his or her intention, to exchange substances with its environment or to carry them to and from bodily parts. In contrast, vitalistic causality indicates that the target phenomenon is caused by activity of an internal organ, which has, like a living thing, "agency" (i.e., a tendency to initiate and sustain behaviors).' The activity is often described as a transmission or exchange of the "vital force",

Young children's naive theory of biology

159

which can be conceptualized as unspecified substance, energy, or information. Vitalistic causality is clearly different from person-intentional causality in the sense that the organ's activities inducing the phenomenon are independent of the intention of the person who possesses the organ. In Inagaki and Hatano (1990) some of the children of ages 5-8 gave explanations referring to something like vital force as a mediator when given novel questions about bodily processes, such as, what the halt of blood circulation would cause; for example, one child said, "If blood does not come to the hands, they will die, because the blood does not carry energies to them"; and another child, "We wouldn't be able to move our hands, because energies fade away if blood does not come there." However, as the number of these children was small, we did another experiment to induce children to choose a plausible explanation out of the presented ones. We (Inagaki & Hatano, 1993, Experiment 2) predicted that even if young children could not apply mechanical causality, and if they could not generate vitalistic causal explanations for themselves, they would prefer vitalistic explanations to intentional ones for bodily processes when asked to choose one from among several possibilities. We asked 6-year-olds, 8-year-olds, and college students as subjects to choose one from three possible explanations each for six biological phenomena, such as blood circulation and breathing. The three explanations represented intentional, vitalistic and mechanical causality, respectively. An example question on blood circulation with three alternative explanations was as follows: "Why do we take in air? (a) Because we want to feel good [intentional]; (b) Because our chest takes in vital power from the air [vitalistic]; (c) Because the lungs take in oxygen and change it into useless carbon dioxide [mechanical]." The 6-year-olds chose vitalistic explanations as most plausible most often; they chose them 54% of the time. With increasing age the subjects came to choose mechanical explanations most often. It should be noted that the 6-year-olds applied non-intentional (vitalistic plus mechanical) causalities 75% of the time, though they were more apt to adopt intentional causality than the 8-year-olds or adults. This vitalistic causality is probably derived from a general mechanism of personification. One who has no means for observing the opaque inside or details of the target object often tries to understand it in a global fashion, by assuming it or its components to be human-like (Ohmori, 1985). Hence, young children try to understand the workings of internal bodily organs by regarding them as humanlike (but non-communicative) agents, and by assigning their activities global life-sustaining characters, which results in vitalistic causality for bodily processes. We can see a similar mode of explanation in the Japanese endogenous science before the Meiji restoration (and the beginning of her rapid modernization),

160

G. Hatano, K. Inagaki

which had evolved with medicine and agriculture as its core (Hatano & Inagaki 1987). ' Young children seem to rely on vitalistic causality only for biological phenomena. They seldom attribute social-psychological behavior, which is optional and not needed for survival, to the agency of a bodily organ or part, as revealed by Inagaki and Hatano (1993, Experiments 3 and 3a). The following is an example question for such behavior used in the study: "When a pretty girl entered the room, Taro came near her. Why did he do so?" Eighty per cent of the 6-year-olds chose (a) "Because Taro wanted to become a friend of hers" [intentional explanation], whereas only 20% opted for (b) "Because Taro's heart urged him to go near her" [vitalistic]. For biological phenomenon questions - almost the same as those used in Experiment 2 of Inagaki and Hatano (1993) except for excluding the mechanical causal explanation - they tended to choose vitalistic explanations rather than intentional ones. What, then, is the relationship between the vitalistic explanation for biological phenomena and the teleological-functional explanation for biological properties (Keil, 1992)? Both are certainly in between the intentional and the mechanical; both seem to afford valid perspectives of the biological world. One interpretation is that they are essentially the same idea with different emphases-the teleological concerns more the why or the cause, whereas the vitalistic is concerned more with the how or the process. Another interpretation is that, because the vitalistic explanation refers to activity of the responsible organ or bodily part (implicitly for sustaining life), it is closer to mechanical causality than is the teleological one, which refers only to the necessity. Anyway, it will be intriguing to examine these characterizations of young children's "biological" explanations in concrete experimental studies. In sum, we can conclude from the above findings that children as young as 6 years of age possess three essential components of biology. In other words, contrary to Carey (1985), children before schooling have acquired a form of biology differentiated from psychology.

Functions of naive biology The personifying and vitalistic biology that young children have is adaptive'in nature in their everyday life. In other words, we believe that naive biology is formed, maintained, and elaborated, because it is functional. First, it is useful in everyday biological problem solving, for example, in making predictions for reactions of familiar animate entities to novel situations, and for properties and behaviors of unfamiliar entities. This is taken for granted, because naive biology includes constrained personification as a general mode of inference for biological phenomena and about biological kinds. Since a human being is a kind of living

Young children's naive theory of biology

161

entity, the use of the person analogy can often produce reasonable, though not necessarily accurate, predictions, especially for advanced animals. The person analogies in young children's biology are especially useful, because they are constrained by the perceived similarity of the target objects to humans and also by specific knowledge regarding the target objects, as revealed by Inagaki and Hatano (1987, 1991). Second, naive biology also enables young children to make sense of biological phenomena they observe in their daily lives. Our favorite example is a 5-year-old girl's statement reported by Motoyoshi (1979). Based on accumulated experience with raising flowers, and relying on her naive biology, she concluded: "Flowers are like people. If flowers eat nothing (are not watered), they will fall down of hunger. If they eat too much (are watered too often), they will be taken ill." In this sense, young children's naive biology constitutes what Keil (1992) calls a mode of construal. Hatano and Inagaki (1991b) presented to 6-year-olds three bodily phenomena of a squirrel, which can also be observed for humans (being constipated, diarrhea, and getting older and weaker), and asked them to guess a cause for each phenomenon. About three-quarters of them on the average could offer some reasonable causes, and also judge the plausibility of causes suggested by the experimenter. About a half of them explicitly referred to humans at least once in their causal attributions for a squirrel. At the same time, however, some of their expressions strongly suggest that they edited or adapted to this animal those responses obtained by the person analogy (e.g., "A squirrel became weaker because it did not eat chestnuts"). Naive biology provides young children with a conceptual tool for interpreting bodily phenomena of other animals as well as humans. Third, naive biology is useful because it helps children learn meaningfully, or even discover, procedures for taking care of animals and plants as well as themselves in everyday life. Global understanding of internal bodily functions is enough for such purposes. Inagaki and Kasetani (1994) examined whether inducing the person analogy, a critical component of naive biological knowledge, would enhance 5- and 6-year-olds' comprehension of raising procedures of a squirrel. The subjects were aurally given the description of the procedures while watching pictures visualizing them. The description included several references to humans in the experimental condition but not in the control condition. For example, about the necessity of giving a variety of food to a squirrel, the experimenter indicated, "You do not eat favorite food only. You eat a variety of food, don't you?" After listening to the description of all procedures, the children were asked to tell how to raise a squirrel to another lady. They were asked questions by this lady, for example, "What kind of food might I give a squirrel? Favorite chestnuts only, chestnuts, seeds and vegetables mixed, or ice cream?" They were thus required to choose an alternative and to give the reason. The experimental group children, irrespective of age, gave more often adequate

162

G. Hatano, K. Inagaki

reasons for their correct choices than the control ones, though their superiority in the number of correct choices was significant only for the younger subjects. For instance, one 5-year-old child said, "Don't feed chestnuts only. You must give a squirrel plenty of seeds and carrots, because a person will die if he eats the same kind of food only, and so will it." An anecdotal but impressive example of the discovery of a procedure for coping with trouble with a raised animal is reported by Motoyoshi (1979). Children aged 5 in a day care center inferred that, when they observed unusual excretion of a rabbit they were taking care of every day, it might be suffering from diarrhea like a person, and after group discussion they produced an idea of making the rabbit take medicine for diarrhea as a suffering person would. Young children's naive biology is functional partly because its componentspieces of knowledge, the mode of inference, and causality - are promptly and effortlessly retrieved and used to generate more or less plausible ideas. Their personifying and vitalistic biology seems to be triggered almost automatically whenever children come into contact with novel phenomena which they recognize as "biological" (Inagaki, 1990b). Speaking generally, making an educated guess by applying insufficient knowledge is often rewarded in everyday life, both in individual problem solving and in social interaction, so most everyday knowledge is readily used. Children's naive biology is not an exception, we believe. In fact in our study described above (Inagaki & Hatano, 1987) it was very rare that the children gave no prediction or the "I don't know" answer to our questions which were somewhat unusual. It should also be noted that naive biological knowledge is seldom applied "mechanically". As mentioned earlier, children constrain their analogies by factual, procedural or conceptual knowledge about the target to generate a reasonable answer.

Acquisition of naive biology As already mentioned, our experimental data strongly suggest that children as young as 6 years of age have acquired a form of biology. This early acquisition of biology is not surprising from the perspective of human evolution, because it has been essential for our species to have some knowledge about animals and plants as potential foods (Wellman & Gelman, 1992) and also knowledge about our bodily functions and health (Hatano, 1989; Inagaki & Hatano, 1993). When children acquire an autonomous domain of biology is still an open question for us, because we have not examined whether much younger subjects too possess a form of biology. However, we think that the acquisition of biology comes a little later than that of physics or psychology. Infants seldom need biological knowledge, since they do not need to take care of their health nor try to find food themselves. Moreover,

Young children's naive theory of biology

163

autonomous biology has to deal with entities which have agency (i.e., initiate and maintain activity without external forces) but can hardly communicate with us humans, and thus has to apply an intermediate form of causality between the intentional and mechanical. Autonomous biology also requires to include animals and plants, which appear so different, into an integrated category of living things. Though there is some evidence that even infants can distinguish objects having a capacity for self-initiated movement from those not having it (e.g., Golinkoff, Harding, Carlson, & Sexton, 1984), this cannot directly serve as the basis for the living-non-living distinction.

Cognitive bases of naive biology Whether naive biology gradually emerges out of naive psychology (Carey, 1985) or is a distinct theory or mode of construal from the start (Keil, 1992) is still debatable. It is true that, as Keil argues, preschool children have some understanding of the distinction between the biological and the social-psychological. In Vera and Keil (1988), for example, 4-year-olds' inductions about animals, when given the biological context, resembled those previously found for 7-year-olds, who were given the same attribution questions without context; giving the social-psychological context to 4-year-olds did not affect the inductions they made. However, young children may overestimate the controllability of bodily processes by will or intention. In fact, our modified replication study on the controllability of internal bodily functions suggests that 3-year-olds are not sure whether the workings of bodily organs are beyond their control (Inagaki & Suzuki, 1991). Our own speculation about how young children acquire personifying and vitalistic biology through everyday life experiences is as follows. Children notice through somatosensation that several "events", uncontrolled by their intention, are going on inside the body. Since children cannot see the inside of the body, they will try to achieve "global understanding" by personifying an organ or bodily part. Considering that young children use analogies in a selective, constrained way (Inagaki & Hatano, 1987, 1991; Vosniadou, 1989), it is plausible that they apply the person analogy to bodily organs in that way, too. More specifically, they attribute agency and some related human properties but not others (e.g., the ability to communicate) to these organs. They also through personification generalize this global understanding of the body to other living things. A set of specific innate or very early cognitive constraints is probably another important factor in the acquisition of naive biology. It is likely that even very young children have tendencies to attribute a specific physical reaction to a specific class of events, such as that diarrhea is caused by eating something poisonous. These tendencies enhance not only their rejection of intentional

164

G. Hatano, K. Inagaki

causality for bodily phenomena but also their construction of more specific beliefs about bodily processes. To sum up, we believe that the ability of young children to make inferences about bodily processes, as well as about animals' and plants' properties and behaviors, is based on personification, but this does not mean it is purely psychological, because they understand the mind-body distinction to some extent. However, it is suggested that they sometimes overattribute human mental properties (not communication ability, but working hard, being happy and others in addition to agency) to bodily organs (Hatano & Inagaki, unpublished study) as well as to less advanced animals, plants or even inanimate objects (e.g., Inagaki & Sugiyama, 1988). In this sense, early naive biology is "psychological", though it is autonomous.

Activity-based experiences We are willing to admit that, because of the above general mechanism of personification and the resultant vitalistic causality, which "fit nicely with biology" (Keil, 1992, p. 105), and specific cognitive constraints, there must be some core elements in naive biology that are shared among individuals within and between cultures, as suggested by Atran (1990). However, we would like to emphasize that this condition does not mean children's activity-based experiences do not contribute to the acquisition. Some such experiences are also universal in human ways of living, but others may vary and thus produce differently instantiated versions of naive biology. For example, if children are actively engaged in raising animals, it will be possible for them to acquire a rich body of knowledge about them, and therefore to use that body of knowledge, as well as their knowledge about humans, as a source for analogical predictions and explanations for other biological kinds. Our studies have in fact revealed that such an activity may produce a slightly different version of naive biology from the ordinary one. Inagaki (1990a) compared the biological knowledge of kindergarteners who had actively engaged in raising goldfish for an extended period at home with that of children of the same age who had never raised any animal. Although these two groups of children did not differ in factual knowledge about typical animals in general, the goldfish-raisers had much richer procedural, factual and conceptual knowledge about goldfish. More interestingly, the goldfish-raisers used the knowledge about goldfish as a source for analogies in predicting reactions of an unfamiliar "aquatic" animal (i.e., a frog), one that they had never raised, and produced reasonable predictions with some explanations for it. For example, one of the raisers answered, when asked whether we could keep a baby frog in the same size forever, "No, we can't, because a frog will grow bigger as goldfish grew bigger.

Young children's naive theory of biology

165

My goldfish were small before, but now they are big." It might be added that the goldfish-raisers tended to use person analogies as well as goldfish analogies for a frog. In other words, the goldfish-raisers could use two sources for making analogical predictions. Moreover, in another study (Kondo & Inagaki, 1991; see also Hatano & Inagaki, 1992), goldfish-raising children tended to enlarge their previously possessed narrow conception of animals. Goldfish-raisers attributed animal properties which are shared by humans (e.g., having a heart, excreting) not only to goldfish but also to a majority of animals phylogenetically between humans and goldfish at a higher rate than non-raisers. This suggests that the experience of raising goldfish modifies young children's preferred mode of biological inferences.

Theory changes in biological understanding So far we have emphasized strengths of young children's naive biology. What weaknesses does it have? Its weaknesses are obvious even when compared with intuitive biology in lay adults. Let us list some major ones: (a) limited factual knowledge; (b) lack of inferences based on complex, hierarchically organized biological categories; (c) lack of mechanical causality; and (d) lack of some conceptual devices (e.g., "evolution", "photosynthesis"). The use of inferences based on complex, hierarchically organized biological categories and of mechanical causality requires a theory change or conceptual change (i.e., fundamental restructuring of knowledge), whereas the accumulation of more and more factual knowledge can be achieved by enrichment only. Whether the acquisition of basic conceptual devices in scientific or school biology is accompanied by a theory change is not beyond dispute, but incorporating them meaningfully into the existing body of knowledge can usually be achieved only with the restructuring of that knowledge. It is expected that, as children grow older, their personifying and vitalistic biology will gradually change toward truly "non-psychological" (if not scientific) biology by eliminating the above weaknesses (b) and (c), that is, toward a biology which relies on category-based inferences and rejects intentional causal explanations. We assume that this change is almost universal, at least among children growing up in highly technological societies, and that it can occur without systematic instruction in biology, though schooling may have some general facilitative effects on it. Inagaki and Sugiyama (1988) examined how young children's human-centered or "similarity-based" inference would change as they grew older. They gave attribution questions, such as "Does X have a property Y?", to children aged from 4 to 10 and college students. Results indicated that there was a progression from 4-year-olds' predominant reliance on similarity-based attribution (attributing

166

G. Hatano, K. Inagaki

human properties in proportion to perceived similarity between target objects and humans) to adults' predominant reliance on category-based attribution (attributing by relying on the higher-order category membership of the targets and category-attribute associations). This shift is induced not only by an increased amount of knowledge but also by the development of metacognitive beliefs evaluating more highly the usefulness of higher-order categories (Hatano & Inagaki, 1991a, Inagaki, 1989). In contrast to young children's vitalistic, and sometimes even intentional, biological explanations, older children reject intentional explanations for biological phenomena and are inclined to use mechanical causality exclusively. In Experiment 2 of Inagaki and Hatano's (1993) study, the difference between 6-year-olds and 8-year-olds was larger than the difference between 8-year-olds and adults in terms of preference for mechanical explanations and avoidance of intentional ones. These results suggest that young children's biology is qualitatively different from the biology that older children and adults have, and that, in accordance with Carey's claim, there occurs a conceptual change in biological understanding between ages 4 and 10. However, contrary to her claim, this change is characterized not as the differentiation of biology from psychology but as a qualitative change within the autonomous domain of biology, because children as young as 6 years of age already possess a form of biology. Another important change may occur as a result of the learning of scientific biology at school. In order to be able to reason "scientifically" in biology one needs to know its basic concepts and principles - major conceptual devices which cannot be acquired without intervention. For example, if one does not know the phenomenon of photosynthesis, one will not be able to understand the difference between animals and plants (i.e., plants can produce nutriment themselves), and thus may accept the false analogy of mapping water for plants with food for animals. We assume that, unlike the first theory change, this change is hard to achieve and thus occurs only among a limited portion of older children or adolescents.

Universality of naive biology Which aspects of naive biology are universal, and which aspects are not? As suggested by Atran (1990), it may be possible to find the "common sense" or core beliefs shared by all forms of folk biology and even by scientific biology. However, what such core beliefs are is debatable. Much of the research inspired by Piaget has shown parallels among the biological understanding of children in different cultures. The distinctions between animals and terrestrial inanimate objects are particularly strong. However,

Young children's naive theory of biology

167

the biological understanding observed in different cultures is not identical. The most striking of the differences thus far reported concerns ideas about plants of children in Israel. Stavy and Wax (1989) showed that about half of a sample of 6-12-year-olds, when asked to judge the life status of animals, plants and non-living things, classified plants either as non-living things or as falling within a third category: things that are neither living nor non-living. Beliefs about inanimate objects also may differ between cultures. Whereas recent studies conducted in North America indicate that young children seldom attribute life or other living-thing properties to any terrestrial inanimate objects (e.g., Dolgin & Behrend, 1984; Richards & Siegler, 1984), Inagaki and Sugiyama (1988) reported that some Japanese preschoolers extended mental properties even to inanimate objects without movement or function, such as stones. Hatano et al. (1993) tried to differentiate between universal and culturally specific aspects of children's conceptions of life and understanding of attributes of living things, by comparing kindergarteners, 2nd- and 4th-graders from Israel, Japan and the United States. The children were asked whether two instances each of four object types (people, other animals, plants and inanimate objects) possessed each of 16 attributes that included life status (being alive), unobservable animal attributes (e.g., has a heart), sensory attributes (e.g., feels pain), and attributes true of all living things (e.g., grows bigger). The results illustrate both similarities and differences across cultures in children's biological understanding. Children in all cultures knew that people, other animals, plants, and inanimate objects were different types of entities, with different properties, and were extremely accurate regarding humans, somewhat less accurate regarding other animals and inanimate objects, and least accurate regarding plants. At the same time, as predicted from cultural analyses, Israeli children were considerably more likely not to attribute to plants properties that are shared by all living things, whereas Japanese children, whose overall accuracy was comparable to the Israeli, were considerably more likely to attribute to inanimate objects properties that are unique to living things. These differences are especially interesting because they suggest that children's naive biology is influenced by beliefs within the culture where they grow up. Consider why Japanese children might be more likely than children in the United States or Israel to view plants or inanimate objects as alive and having attributes of living things. Japanese culture includes a belief that plants are much like human beings. This attitude is represented by the Buddhist idea that even a tree or blade of grass has a mind. In Japanese folk psychology, even inanimate objects are sometimes considered to have minds. For example, it is at least not a silly idea for Japanese to assign life or divinity not only to plants but also to inanimate objects, especially big or old ones. In addition, linguistic factors seem to influence Japanese children's attributional judgements. The kanji (Chinese character) representing it has a prototypal meaning of "fresh" or "perishable" as well as

168

G. Hatano, K. Inagaki

"alive". Therefore, this kanji can be applied to cake, wine, sauce, and other perishable goods. Similar features of culture and language may account for Israeli children being less apt than American or Japanese children to attribute to plants life status and properties of living things. Stavy and Wax (1989) suggested that within the Israeli culture plants are regarded as very different from humans and other animals in their life status. This cultural attitude parallels that of a biblical passage (Genesis 1: 30), well known to Israeli students, indicating that plants were created as food for living things including animals, birds and insects. Adding to, or perhaps reflecting, their cultural beliefs, the Hebrew word for "animal" is very close to that for "living" and"alive". In contrast the word for "plant" has no obvious relation to such terms (Stavy & Wax, 1989). How culture influences the development of biological understanding has yet to be studied. Parents, schools and mass media may serve to transmit cultural beliefs. For example, Japanese parents may communicate the attitude through their actions toward plants and divine inanimate objects, though they do not usually tell their children this explicitly. Culture may provide children with opportunities to engage in activities that lead them to construct some particular biological understanding, as in the case of children raising goldfish (Hatano & Inagaki, 1992; Inagaki, 1990a).

Postscript Since Carey (1985), young children's naive biology has been an exciting topic for research in cognitive development. As more and more ambitious researchers have joined to study it, not only has a richer database been built and finer conceptualizations offered about this specific issue, but also, through attempts to answer questions like the ones discussed so far in this article, a better understanding of fundamental issues in the developmental studies on cognition, like the nature of domains, theories, constraints, etc., has been achieved. It will probably be a popular topic for the coming several years, and research questions about naive biology can be better answered and/or better rephrased. What is urgently needed now is (a) to integrate nativistic and cultural accounts of acquisition and change in naive biology, and (b) to find commonalities and differences between naive biology and other major theories of the world possessed by young children (Hatano, 1990).

References Atran, S. (1990). Cognitive foundations of natural history: Towards an anthropology of science. Cambridge, UK: Cambridge University Press.

Young children's naive theory of biology

169

Backscheider, A.G., Shatz, M., & Gelman, S.A. (1993). Preschoolers' ability to distinguish living kinds as a function of regrowth. Child Development, 64, 1242-1257. Carey, S. (1985). Conceptual change in childhood. Cambridge, MA: MIT Press. Carey, S. (1987). Theory change in childhood. In B. Inhelder, D. de Caprona, & A. Cornu-Wells (Eds.), Piaget today (pp. 141-163). Hillsdale, NJ: Erlbaum. Dolgin, K.G., & Behrend, D.A. (1984). Children's knowledge about animates and inanimates. Child Development, 55, 1646-1650. Furth, H.G. (1980). The world of grown-ups: Children's conceptions of society. New York: Elsevier. Gellert, E. (1962). Children's conceptions of the content and functions of the human body. Genetic Psychology Monographs, 65, 291-411. Gelman, R. (1990). First principles organize attention to and learning about relevant data: Number and the animate-inanimate distinction as examples. Cognitive Science, 14, 79-106. Golinkoff, R.M., Harding, C.G., Carlson, V., & Sexton, M.E. (1984). The infant's perception of causal events: The distinction between animate and inanimate objects. In L.L. Lipsitt & C. Rovee-Collier (Eds.), Advances in Infancy Research (Vol. 3, pp. 145-165). Norwood, NJ: Ablex. Hatano, G. (1989). Language is not the only universal knowledge system: A view from "everyday cognition". Dokkyo Studies in Data Processing and Computer Science, 7, 69-76. Hatano, G. (1990). The nature of everyday science: A brief introduction. British Journal of Developmental Psychology, 8, 245-250. Hatano, G., & Inagaki, K. (1987). Everyday biology and school biology: How do they interact? Quarterly Newsletter of the Laboratory of Comparative Human Cognition, 9, 120-128. Hatano, G., & Inagaki, K. (1991a). Learning to trust higher-order categories in biology instruction. Paper presented at the meeting of the American Educational Research Association, Chicago. Hatano, G., & Inagaki, K. (1991b). Young children's causal reasoning through spontaneous personification. Paper presented at the 33rd meeting of the Japanese Educational Psychology Association, Nagano [in Japanese]. Hatano, G., & Inagaki, K. (1992). Desituating cognition through the construction of conceptual knowledge. In P. Light & G. Butterworth (Eds.), Context and cognition: Ways of learning and knowing (pp. 115-133). London: Harvester/Wheatsheaf. Hatano, G., & Inagaki, K. (1994). Recognizing commonalities between animals and plants. Paper to be presented at the meeting of the American Educational Research Association, New Orleans. Hatano, G., Siegler, R.S., Richards, D.D., Inagaki, K., Stavy, R., & Wax, N. (1993). The development of biological knowledge: A multi-national study. Cognitive Development, 8, 47-62. Inagaki, K. (1989). Developmental shift in biological inference processes: From similarity-based to category-based attribution. Human Development, 32, 79-87. Inagaki, K. (1990a). The effects of raising animals on children's biological knowledge. British Journal of Developmental Psychology, 8, 119-129. Inagaki, K. (1990b). Young children's use of knowledge in everyday biology. British Journal of Developmental Psychology, 8, 281-288. Inagaki, K. (1993a). Young children's differentiation of plants from non-living things in terms of growth. Paper presented at the 60th meeting of the Society for Research in Child Development, New Orleans. Inagaki, K. (1993b). The nature of young children's naive biology. Paper presented at the symposium, "Children's naive theories of the world", at the 12th meeting of the International Society for the Study of Behavioral Development, Recife, Brazil. Inagaki, K., & Hatano, G. (1987). Young children's spontaneous personification as analogy. Child Development, 58, 1013-1020. Inagaki, K., & Hatano, G. (1990). Development of explanations for bodily functions. Paper presented at the 32nd meeting of the Japanese Educational Psychology Association, Osaka [in Japanese]. Inagaki, K., & Hatano, G. (1991). Constrained person analogy in young children's biological inference. Cognitive Development, 6, 219-231. Inagaki, K., & Hatano, G. (1993). Young children's understanding of the mind-body distinction. Child Development, 64, 1534-1549. Inagaki, K., & Kasetani, M. (1994). Effects of hints to use knowledge about humans on young

170

G. Hatano, K. Inagaki

children's understanding of biological phenomena. Paper to be presented at the 13th meeting of the International Society for the Study of Behavioral Development, Amsterdam. Inagaki, K., & Sugiyama, K. (1988). Attributing human characteristics: Developmental changes in over- and underattribution. Cognitive Development, 3, 55-70. Inagaki, K., & Suzuki, Y. (1991). The understanding of the mind-body distinction in children aged 3 to 5 years. Paper presented at the 33rd meeting of the Japanese Educational Psychology Association, Nagano, [in Japanese]. Keil, F.C. (1992). The origins of an autonomous biology. In M.R. Gunnar & M. Maratsos (Eds.), Modularity and constraints in language and cognition. Minnesota Symposia on Child Psychology (Vol. 25, pp. 103-137). Hillsdale, NJ: Erlbaum. Kondo, H., & Inagaki, K. (1991). Effects of raising goldfish on the grasp of common characteristics of animals. Paper presented at the 44th Annual Meeting of Japanese Early Childhood Education and Care Association, Kobe [in Japanese]. Kuhn, D. (1989). Children and adults as intuitive scientists. Psychological Review, 96, 674-689. Massey, CM., & Gelman, R. (1988). Preschooler's ability to decide whether a photographed unfamiliar object can move itself. Developmental Psychology, 24, 307-317. Motoyoshi, M. (1979). Watashino Seikatuhoikuron. [Essays on education for day care children: Emphasizing daily life activities.] Tokyo: Froebel-kan [in Japanese]. Ohmori, S. (1985). Chishikito gakumonno kouzou. [The structure of knowledge and science.] Tokyo: Nihon Hoso Shuppan Kyokai [in Japanese]. Richards, D.D., & Siegler, R.S. (1984). The effects of task requirements on children's life judgments. Child Development, 55, 1687-1696. Rosengren, K.S., Gelman, S.A., Kalish, C.W., & McCormick, M. (1991). As time goes by: Children's early understanding of growth. Child Development, 62, 1302-1320. Siegal, M. (1988). Children's knowledge of contagion and contamination as causes of illness. Child Development, 59, 1353-1359. Smith, C, Carey, S., & Wiser, M. (1985). On differentiation: A case study of the development of the concepts of size, weight, and density. Cognition, 21, 177-237. Springer, K., & Keil, F.C. (1989). On the development of biologically specific beliefe: The case of inheritance. Child Development, 60, 637-648. Stavy, R., & Wax, N. (1989). Children's conceptions of plants as living things. Human Development, 32, 88-94. Vera, A.H., & Keil, F.C. (1988). The development of inductions about biological kinds: The nature of the conceptual base. Paper presented at the 29th meeting of the Psychonomic Society, Chicago. Vosniadou, S. (1989). Analogical reasoning as a mechanism in knowledge acquisition: A developmental perspective. In S. Vosniadou & A. Ortony (Eds.), Similarity and analogical reasoning (pp. 413-469). Cambridge, UK: Cambridge University Press. Vosniadou, S., & Brewer, W. (1992). Mental models of the earth: A study of conceptual change in childhood. Cognitive Psychology, 24 535-585. Wellman, H.M. (1990). The child's theory of mind. Cambridge, MA: MIT Press. Wellman, H.M., & Gelman, S.A. (1992). Cognitive development: Foundational theories of core domains. Annual Review of Psychology, 43, 337-375.

9 Mental models and probabilistic thinking Philip N. Johnson-Laird* Department of Psychology, Princeton University, Green Hall, Princeton, NJ 08544, USA

Abstract This paper outlines the theory of reasoning based on mental models, and then shows how this theory might be extended to deal with probabilistic thinking. The same explanatory framework accommodates deduction and induction: there are both deductive and inductive inferences that yield probabilistic conclusions. The framework yields a theoretical conception of strength of inference, that is, a theory of what the strength of an inference is objectively: it equals the proportion of possible states of affairs consistent with the premises in which the conclusion is true, that is, the probability that the conclusion is true given that the premises are true. Since there are infinitely many possible states of affairs consistent with any set of premises, the paper then characterizes how individuals estimate the strength of an argument. They construct mental models, which each correspond to an infinite set of possibilities (or, in some cases, a finite set of infinite sets of possibilities). The construction of models is guided by knowledge and beliefs, including lay conceptions of such matters as the i(law of large numbers'9. The paper illustrates how this theory can account for phenomena of probabilistic reasoning.

1. Introduction Everyone from Aristotle to aboriginals engages in probabilistic thinking, whether or not they know anything of the probability calculus. Someone tells you: *Fax (609) 258 1113, e-mail [email protected] The author is grateful to the James S. McDonnell Foundation for support. He thanks Jacques Mehler for soliciting this paper (and for all his work on 50 volumes of Cognition]). He also thanks Ruth Byrne for her help in developing the model theory of deduction, Eldar Shafir for many friendly discussions and arguments about the fundamental nature of probabilistic thinking, and for his critique of the present paper. Malcolm Bauer, Jonathan Evans and Alan Garnham also kindly criticized the paper. All these individuals have tried to correct the erroneous thoughts it embodies. Thanks also to many friends - too numerous to mention - for their work on mental models.

172

P. Johnson-Laird

There was a severe frost last night. and you are likely to infer: The vines will probably not have survived it. basing the inference on your knowledge of the effects of frost. These inferences are typical and ubiquitous. They are part of a universal human competence, which does not necessarily depend on any overt mastery of numbers or quantitative measures. Aristotle's notion of probability, for instance, amounts to the following two ideas: a probability is a thing that happens for the most part, and conclusions that state what is probable must be drawn from premises that do the same (see Rhetoric, I, 1357a). Such ideas are crude in comparison to Pascal's conception of probability, but they correspond to the level of competence a psychological theory should initially aspire to explain. Of course many people do encounter the probability calculus at school. Few master it, as a simple test with adults shows: There are two events, which each have a probability of a half. What is the probability that both occur? Many people respond: a quarter. The appropriate "therapy" for such errors is to invite the individual first to imagine that A is a coin landing heads and B is the same coin landing tails, that is, p(A & B) = 0, and then to imagine that A is a coin landing heads and B is a coin landing with the date uppermost, where date and head are on the same side, that is, p(A & B) = 0.5. At this point, most people begin to grasp that there is no definite answer to the question above - joint probabilities are a function of the dependence of one event on the other. Cognitive psychologists have discovered many phenomena of probabilistic thinking, principally that individuals do not follow the propositional calculus in assessing probabilities, and that they appear to rely on a variety of heuristics in making judgements about probabilities. A classic demonstration is Tversky and Kahneman's (1983) phenomenon of the "conjunction fallacy", that is, a violation of the elementary principle that p(A & B)^p(B). For example, subjects judge that a woman who is described as 31 years old, liberal and outspoken, is more likely to be a feminist bankteller than a bankteller. Indeed, we are all likely to go wrong in thinking about probabilities: the calculus is a branch of mathematics that few people completely master. Theorists relate probability to induction, and they talk of both inductive inference and inductive argument. The two expressions bring out the point that the informal arguments of everyday life, which occur in conversation, newspaper

Mental models and probabilistic thinking

173

editorials and scientific papers, are often based on inductive inferences. The strength of such arguments depends on the relation between the premises and the conclusion. But the nature of this relation is deeply puzzling - so puzzling that many theorists have abandoned logic altogether in favor of other idiosyncratic methods of assessing informal arguments (see, for example, Toulmin, 1958; the movement for "informal logic and critical thinking", e.g. Fisher, 1988; and "neural net" models, e.g. Thagard, 1989). Cognitive psychologists do not know how people make probabilistic inferences: they have yet to develop a computable account of the mental processes underlying such reasoning. For this celebratory volume of Cognition, the editor solicited papers summarizing their author's contributions to the field. The present paper, however, looks forward as much as it looks back. Its aim is to show how probabilistic thinking could be based on mental models-an approach that is unlikely to surprise assiduous readers of the journal (see, for example, Byrne, 1989; Johnson-Laird & Bara, 1984; Oakhill, Johnson-Laird, & Garnham, 1989). In pursuing the editor's instructions, part 2 of the paper reviews the theory of mental models in a self-contained way. Part 3 outlines a theoretical conception of strength of inference, that is, a theory of what objectively the strength of an inference or argument depends on. This abstract account provides the agenda for what the mind attempts to compute in thinking probabilistically ( a theory at the "computational" level; Marr, 1982). However, as we shall see, it is impossible for a finite device, such as the human brain, to carry out a direct assessment of the strength of an inference except in certain limiting cases. Part 4 accordingly describes a theory of how the mind attempts to estimate the strength of inferences (a theory at the "algorithmic" level). Part 5 shows how this algorithmic theory accounts for phenomena of probabilistic thinking and how it relates to the heuristic approach. Part 6 contrasts the model approach with theories based on rules of inference, and shows how one conception of rules can be reconciled with mental models.

2. Reasoning and mental models Mental models were originally proposed as a programmatic basis for thinking (Craik, 1943). More recently, the theory was developed to account for verbal comprehension: understanding of discourse leads to a model of the situation under discussion, that is, a representation akin to the result of perceiving or imagining the situation. Such models are derived from syntactically structured expressions in a mental language, which are constructed as sentences are parsed (see Garnham, 1987; Johnson-Laird, 1983). Among the key properties of models is that their structure corresponds to the structure of what they represent (like a visual image), and thus that individual entities are represented just once in a model. The theory of mental models has also been developed to explain deductive

174

P. Johnson-Laird

reasoning (Johnson-Laird, 1983; Johnson-Laird & Byrne, 1991). Here, the underlying idea is that reasoning depends on constructing a model (or set of models) based on the premises and general knowledge, formulating a conclusion that is true in the model(s) and that makes explicit something only implicit in the premises, and then checking the validity of the conclusion by searching for alternative models of the premises in which it is false. If there are no such counterexamples, then the conclusion is deductively valid, that is, it must be true given that the premises are true. Thus, the first stage of deduction corresponds to the normal process of verbal comprehension, the second stage corresponds to the normal process of formulating a useful and parsimonious description, and only the third stage is peculiar to reasoning. To characterize any particular domain of deduction, for example reasoning based on temporal relations such as "before", "after" and "while", or sentential connectives such as "not", "if, "and" and "or", it is necessary to account for how the meanings of the relevant terms give rise to models. The general reasoning principles, as outlined above, then automatically apply to the domain. In fact, the appropriate semantics has been outlined for temporal relations, spatial relations, sentential connectives and quantifiers (such as "all", "none" and "some"), and all of these domains can be handled according to five representational principles: (1) Each entity is represented by an individual token in a model, its properties are represented by properties of the token, and the relations between entities are represented by the relations between tokens. Thus, a model of the assertion "The circle is on the right of the triangle" has the following spatial structure: A

O

which may be experienced as a visual image, though what matters is not so much the subjective experience as the structure of the model. To the extent that individuals grasp the truth conditions of propositions containing abstract concepts, such as friendship, ownership and justice, they must be able to envisage situations that satisfy them, that is, to form mental models of these situations (see JohnsonLaird, 1983, Ch. 15). (2) Alternative possibilities can be represented by alternative models. Thus, the assertion "Either there is a triangle or there is a circle, but not both" requires two alternative models, which each correspond to separate possibilities: A O

(3) The negation of atomic propositions can be represented by a propositional annotation. Thus, the assertion "There is not a triangle" is represented by the following sort of model:

Mental models and probabilistic thinking

175

^A

where "-1" is an annotation standing for negation (for a defence of such annotations, see Polk & Newell, 1988; and Johnson-Laird & Byrne, 1991, pp. 130-1). Of course, the nature of the mental symbol corresponding to negation is unknown. The principal purpose of the annotation is to ensure that models are not formed containing both an element and its negation. Thus, the only way to combine the disjunctive models above with the model of 'There is not a triangle" is to eliminate the first model, leaving only the second model and its new negated element: ^A

O

It follows that there is a circle. As this example shows, deductions can be made without the need for formal rules of inference of the sort postulated in "natural deduction systems" (see, for example, Rips, 1983; Braine, Reiser & Rumain, 1984), such as, in this case, the formal rule for disjunction: A or B not A .-. B (4) Information can be represented implicitly in order to reduce the load on working memory. An explicit representation makes information immediately available to other processes, whereas an implicit information encodes the information in a way that is not immediately accessible. Individuals and situations are represented implicitly by a propositional annotation that works in concert with an annotation for what has been represented exhaustively. Thus, the proper initial representation of the disjunction "Either there is a triangle or there is a circle, but not both" indicates that for the cases in which triangles occur, and the cases in which circles occur, have been exhaustively represented, as shown by the square brackets: [A] [O]

This set of models implicitly represents the fact that circles cannot occur in the first model and triangles cannot occur in the second model, because circles are exhaustively represented in the second model and triangles are exhaustively represented in the first model. Thus, a completely explicit set of models can be constructed by fleshing out the initial models to produce the set: A

-iO

-iA

O

176

P. Johnson-Laird

where there is no longer any need for square brackets because all the elements in the models have been exhaustively represented. The key to understanding implicit information is accordingly the process of fleshing out models explicitly, which is governed by two principles: first, when an element has been exhaustively represented (as shown by square brackets) in one or more models, add its negation to any other models; second, when a proposition has not been exhaustively represented, then add both it and its negation to separate models formed by fleshing out any model in which it does not occur. (Only the first principle is needed to flesh out the models of the disjunction above.) (5) The epistemic status of a model can be represented by a propositional annotation; for example, a model represents a real possibility, a counterfactual state of affairs, or a deontic state. A model that does not contain propositional annotations, that is, a model based on the first two assumptions above, represents a set of possible states of affairs, which contains an infinite number of possibilities (Barwise, 1993). Hence, the model above of the assertion 'The circle is on the right of the triangle" corresponds to infinitely many possibilities; for example, the model is not specific about the distance apart of the two shapes. Any potential counterexample to a conclusion must be consistent with the premises, but the model itself does not enable the premises to be uniquely reconstructed. Hence, in verbal reasoning, there must be an independent record of the premises, which is assumed to be the linguistic representation from which the models are constructed. This record also allows the inferential system to ascertain just which aspects of the world the model represents; for example, a given model may, or may not, represent the distances apart of objects, but inspection of the model alone does not determine whether it represents distance. Experimental evidence bears out the psychological reality of both linguistic representations and mental models (see Johnson-Laird, 1983). Models with propositional annotations compress sets of states of affairs in a still more powerful way: a single model now represents a finite set of alternative sets of situations. This aspect of mental models plays a crucial role in the account of syllogistic reasoning and reasoning with multiple quantifiers. For example, syllogistic premises of the form: All the A are B All the B are C call for one model in which the number of As is small but arbitrary: [[a] [[a]

b] b]

c c

Mental models and probabilistic thinking

177

As are exhaustively represented in relation to Bs, Bs are exhaustively represented in relation to Cs, Cs are not exhaustively related, and the three dots designate implicit individuals of some other sort. This single model supports the conclusion: All the A are C and there are no counterexamples. The initial model, however, corresponds to eight distinct sets of possibilities depending on how the implicit individuals are fleshed out explicitly. There may, or may not, be individuals of each of the three following sorts: individuals who are not-a, not-b, not-c individuals who are not-a, not-b but c individuals who are not-a, but b and c These three binary contrasts accordingly yield eight alternatives, and each of them is consistent with an indefinite number of possibilities depending on the actual numbers of individuals of the different sorts (see also Garnham, 1993). In short, eight distinct potentially infinite sets have been compressed into a single model, which is used for the inference. The theory of reasoning based on mental models makes three principal predictions. First, the greater the number of models that an inference calls for, the harder the task will be. This prediction calls for a theoretical account of the models postulated for a particular domain. Such accounts typically depend on independently motivated psycholinguistic principles; for example, negative assertions bring to mind the affirmative propositions that are denied (Wason, 1965). Second, erroneous conclusions will tend to be consistent with the premises rather than inconsistent with them. Reasoners will err because they construct some of the models of the premises - typically, just one model of them - and overlook other possible models. This prediction can be tested without knowing the detailed models postulated by the theory: it is necessary only to determine whether or not erroneous conclusions are consistent with the premises. Third, knowledge can influence the process of deductive reasoning: subjects will search more assiduously for alternative models when a putative conclusion is unbelievable than when it is believable. The first two of these predictions have been corroborated experimentally for all the main domains of deduction (for a review, see JohnsonLaird & Byrne, 1991, and for a reply to commentators, see Johnson-Laird & Byrne, 1993). The third prediction has been corroborated in the only domain in which it has so far been tested, namely, syllogistic reasoning (see, for example, Oakhill, Johnson-Laird, & Garnham, 1989). In contrast, theories of deduction

178

P. Johnson-Laird

based on formal rules of inference exist only for spatial reasoning and reasoning based on sentential connectives (e.g., Rips, 1983; Braine, Reiser, & Rumain, 1984). Where the model theory and the formal rule theories make opposing predictions, the evidence so far has corroborated the model theory.

3. The strength of an inference By definition, inductive arguments are logically invalid; that is, their premises could be true but their conclusions false. Yet such arguments differ in their strength - some are highly convincing, others are not. These differences are an important clue to the psychology of inference. However, one needs to distinguish between the strength of an argument - the degree to which its premises, if true, support the conclusion, and the degree to which the conclusion is likely to be true in any case. An argument can be strong but its conclusion improbable because the argument is based on improbable premises. Hence, the probability of the premises is distinct from the strength of the argument. In principle, the probability of a conclusion should depend on both the probability of the premises and the strength of the argument. But, as we shall see, individuals are liable to neglect the second of these components. Osherson, Smith, and Shafir (1986) in a ground-breaking analysis of induction explored a variety of accounts of inferential strength that boil down to three main hypotheses: (1) an inference is strong if, given an implicit assumption, schema or causal scenario, it is logically valid; that is, the inference is an enthymeme (cf. Aristotle); (2) an inference is strong if it corresponds to a deduction in reverse, such as argument from specific facts to a generalization of them (cf. Hempel, 1965); and (3) an inference is strong if the predicates (or arguments) in premises and conclusion are similar (cf. Kahneman & Tversky, 1972). Each hypothesis has it advantages and disadvantages, but their strong points can be captured in the following analysis, which we will develop in two stages. First, the present section of the paper will specify an abstract characterization of the objective strength of an argument - what in theory has to be computed in order to determine the strength of an inference (the theory at the "computational" level). Second, the next section of the paper will specify how in practice the mind attempts to assess the strength of an argument (the theory at the "algorithmic" level). The relation between premises and conclusion in inductive inference is a semantic one, and it can be characterized abstractly by adopting the semantic approach to logic (see, for example, Barwise & Etchemendy, 1989). An assertion such as "The circle is on the right of the triangle" is, as we have seen, true in infinitely many different situations; that is, the distance apart of the two shapes can differ, as can their respective sizes, shapes, textures and so on. Yet in all of

Mental models and probabilistic thinking

179

these different states the circle is on the right of the triangle. Philosophers sometimes refer to these different states as "possible worlds" and argue that an assertion is true in infinitely many possible worlds. We leave to one side the issue of whether or not possible worlds are countably infinite. The underlying theory has led to a powerful, though controversial, account of the semantics of natural language (see, for example, Montague, 1974). Armed with the notion of possible states of affairs, we can define the notion of the strength of an inference in the following terms: a set of premises, including implicit premises provided by general and contextual knowledge, lend strength to a conclusion according to two principles: (1) The conclusion is true in at least one of the possible states of affairs in which the premises are true; that is, the conclusion is at least consistent with the premises. If there is no such state of affairs, then the conclusion is inconsistent with the premises: the inference has no strength whatsoever, and indeed there is valid argument in favor of the negation of the conclusion. (2) Possible states of affairs in which the premises are true but the conclusion false (i.e., counterexamples) weaken the argument. If there are no counterexamples, then the argument is maximally strong- the conclusion follows validly from the premises. If there are counterexamples, then the strength of the argument equals the proportion of states of affairs consistent with the premises in which the conclusion is also true. This account has a number of advantages. First, it embraces deduction and induction within the same framework. What underlies deduction is the semantic principle of validity: an argument is valid if its conclusion is true in any state of affairs in which its premises are true. An induction increases semantic information and so its conclusion must be false in possible cases in which its premises are true. Hence, inductions are reverse deductions, but they are the reverse of deductions that throw semantic information away. Second, the probability of any one distinct possible state of affairs (possible world) is infinitesimal, and so it is reasonable to assume that possible states of affairs are close to equi-possible. It follows that a method of integrating the area of a subset of states of affairs provides an extensional foundation for probabilities. The strength of an inference is accordingly equivalent to the probability of the conclusion given the premises. It is 1 in the case of a valid deduction, 0 in the case of a conclusion that is inconsistent with the premises, and an intermediate value for inductions. The two abstract principles, however, are not equivalent to the probability calculus: as we shall see, the human inferential system can attempt to assess the relevant proportions without necessarily using the probability calculus. Likewise, the principles have no strong implications for the correct interpretation of probability, which is a matter for self-conscious philosophical reflection. The

180

P. Johnson-Laird

principles are compatible with interpretations in terms of actuarial frequencies of events, equi-possibilities based on physical symmetry, and subjective degrees of belief (cf. Ramsay, 1926; Hintikka's, 1962, analysis of beliefs in terms of possibility; and for an alternative conception, see Shafer & Tversky's, 1985, discussion of "belief functions"). Hence, an argument (or a probability) may concern either a set of events or a unique event. Individuals who are innumerate may not assign a numerical degree of certainty to their conclusion, and even numerate individuals may not have a tacit mental number representing their degree of belief. Individuals' beliefs do differ in subjective strength, but it does not follow that such differences call for a mental representation of numerical probabilities. An alternative conception of "degrees of belief might be based on analogue representations (cf. Hintzman, Nozawa, & Irmscher, 1982), or on a system that permitted only partial rankings of strengths, such as one that recorded the relative ease of constructing different classes of models. Third, the account is compatible with semantic information. The semantic information conveyed by a proposition, A, equals 1 -p(A)> where p(A) denotes the probability of A (Bar-Hillel & Carnap, 1964; Johnson-Laird, 1983). If A is complex proposition containing conjunctions, disjunctions, etc., its probability can be computed in the usual way according to the probability calculus. Hence, as argued elsewhere (Johnson-Laird, 1993), we can distinguish between deduction and induction on the basis of semantic information, that is, the possible states of affairs that a proposition rules out as false. Deduction does not increase semantic information; that is, the conclusion of a valid deduction rules out the same possibilities as the premises or else fewer possibilities, and so the conclusion must be true given that the premises are true. Induction increases semantic information; that is, the conclusion of an induction goes beyond the premises (including those tacit premises provided by general knowledge) by ruling out at least some additional possibility over and above the states of affairs that they rule out. This account captures all the standard cases of induction, such as the generalization from a finite set of observations to a universal claim (for a similar view, see Ramsay, 1926). Fourth, the account is compatible with everyday reasoning and argumentation. One feature of such informal argumentation is that it typically introduces both a case for a conclusion and a case against it - a procedure that is so unlike a logical proof that many theorists have supposed that logic is useless in the analysis of everyday reasoning (e.g, Toulmin, 1958). The strength of an argument, however, can be straightforwardly analyzed in the terms described above: informal argumentation is typically a species of induction, which may veer at one end into deduction and at the other end into a creative process in which one or more premises are abandoned. Thus, a case for a conclusion may depend on several inductive arguments of differing strength. The obvious disadvantage of the account is that it is completely impractical. No

Mental models and probabilistic thinking

181

one can consider all the infinitely many states of affairs consistent with a set of premises. No one can integrate all those states of affairs in which the conclusion is true and all those states of affairs in which it is false. Inference with quantifiers has no general decision procedure; that is, proofs for valid theorems can always be found in principle, but demonstrations of invalidity may get lost in the uspace" of possible derivations. Inference with sentential connectives has a decision procedure, but the formulation of parsimonious conclusions that maintain semantic information is not computationally tractable; that is, as premises contain more atomic propositions, it takes exponentially longer to generate such conclusions (given that NP ^ P). So how does this account translate into a psychological mechanism for assessing the strength of an argument? It is this problem that the theory of mental models is designed to solve.

4. Mental models and estimates of inferential strength Philosophers have tried to relate probability and induction at a deep level (see, for example, Carnap, 1950), but as far as cognitive psychology is concerned they are overlapping rather than identical enterprises: there are probabilistic inferences that are not inductive, and there are inductive inferences that are not probabilistic. Here, for example, is a piece of probabilistic reasoning that is deductive: The probability of heads is 0.5. The probability of the date uppermost given heads is 1. The probability of the date uppermost given tails is 0. Hence, the probability of the date uppermost is 0.5. This deduction makes explicit what is implicit in the premises, and it does not increase their semantic information. A more mundane example is as follows: If you park illegally within the walls of Siena, you will probably have your car towed. Phil has parked illegally within the walls of Siena. Phil will probably have his car towed. This inference is also a valid deduction. Conversely, many inductive inferences are not probabilistic; that is, they lead to conclusions that people hold to be valid. For example, the engineers in charge at Chernobyl inferred initially that the explosion had not destroyed the reactor (Medvedev, 1990). Such an event was unthinkable from their previous experience, and they had no evidence to suppose that it had occurred. They were certain that the reactor was intact, and their

P. Johnson-Laird

182

conviction was one of the factors that led to the delay in evacuating the inhabitants of the nearby town. Of course people do make probabilistic inductions, and it is necessary to explain their basis as well as the basis for probabilistic deductions. To understand the application of the model theory to the assessment of strength, it will be helpful to consider first how it accounts for deductions based on probabilities. Critics sometimes claim that models can be used only to represent alternative states of affairs that are treated as equally likely. In fact, there is no reason to suppose that when individuals construct or compare models they take each model to be equally likely. To illustrate the point, consider an example of a deduction leading to a probabilistic conclusion: Kropotkin is an anarchist. Most anarchists are bourgeois. .-. Probably, Kropotkin is bourgeois. The quantifier "most" calls for a model that represents a proportion (see Johnson-Laird, 1983, p. 137). Thus, a model of the second premise takes the form: [a] [a] [a] [a]

b b b

where the set of anarchists is exhaustively represented; that is, anarchists cannot occur in fleshing out the implicit model designated by the three dots. When the information in the first premise is added to this model, one possible model is: k

[a] [a] [a] [a]

b b b

in which Kropotkin is bourgeois. Another possible model is:

k

[a] [a] [a] [a]

b b b

Mental models and probabilistic thinking

183

in which Kropotkin is not bourgeois. Following Aristotle, assertions of the form: probably S, can be treated as equivalent to: in most possible states of affairs, S. And in most possible states of affairs as assessed from models of the premises, Kropotkin is bourgeois. Hence, the inferential system needs to keep track of the relative frequency with which the two sorts of models occur. It will detect the greater frequency of models in which it Kropotkin is bourgeois, and so it will deduce: Probably, Kropotkin is bourgeois. Individuals who are capable of one-to-one mappings but who have no access to cardinal or ordinal numbers will still be able to make this inference. They have merely to map each model in which S occurs one-to-one with each model in which S does not occur, and, if there is a residue, it corresponds to the more probable category. Likewise, there are many ways in principle in which to estimate the relative frequencies of the two sorts of model-from random sampling with replacement to systematic explorations of the "space" of possible models. The only difference in induction is that information that goes beyond the premises (including those in tacit knowledge) is added to models on the basis of various constraints (see Johnson-Laird, 1983). The strength of an inference depends, as we have seen, on the relative proportions of two sorts of possible states of affairs consistent with the premises: those in which the conclusion is true and those in which it is false. Reasoners can estimate these proportions by constructing models of the premises and attending to the proportions with which the two sorts of models come to mind, and perhaps to the relative ease of constructing them. For example, given that Evelyn fell (without a parachute) from an airplane flying at a height of 2000 feet, then most individuals have a prior knowledge that Evelyn is likely to be killed, but naive individuals who encounter such a case for the first time can infer the conclusion. The inference is strong, but not irrefutable. They may be able to imagine cases to the contrary; for example, Evelyn falls into a large haystack, or a deep snow drift. But, in constructing models (of sets of possibilities), those in which Evelyn is killed will occur much more often than those in which Evelyn survives - just as models in which Kropotkin is bourgeois outnumber those in which he is not. Insofar as individuals share available knowledge, their assessments of probabilities should be consistent. This account is compatible with the idea of estimating likelihoods in terms of scenarios, which was proposed by Tversky and Kahneman (1973, p. 229), and it forms a bridge between the model theory and the heuristic approach to judgements of probability. Estimates of the relative proportions of the two sorts of models - those in which a conclusion is true and those in which it is false - will

184

P. Johnson-Laird

be rudimentary, biased and governed by heuristics. In assessing outcomes dependent on sequences of events, models must allow for alternative courses of events. They then resemble so-called "event trees", which Shafer (1993) argues provide a philosophical foundation to probability and its relations to causality. Disjunctive alternatives, however, are a source of difficulty both in deduction (see, for example, Johnson-Laird & Byrne, 1991) and in choice (see, for example, Shafir & Tversky, 1992).

5. Some empirical consequences of the theory The strength of an argument depends on the relation between the premises and the conclusion, and, in particular, on the proportion of possibilities compatible with the premises in which the conclusion is true. This relation is not in general a formal or syntactic one, but a semantic one. It takes work to estimate the strength of relation, and the theory yields a number of predictions about making and assessing inductive inferences. The main predictions of the theory are as follows: First, arguments - especially in daily life - do not wear their logical status on their sleeves, and so individuals will tend to approach deductive and inductive arguments alike. They will tend to confuse an inductive conclusion, that is, one that could be true given the premises, with a deductive conclusion, that is, one that must be true given the premises. They will tend to construct one or two models, draw a conclusion, and be uncertain about whether it follows of necessity. Second, envisioning models, which each correspond to a class of possibilities, is a crude method, and, because of the limited processing capacity of working memory, many models are likely never to be envisaged at all. The process will be affected by several constraints. In particular, individuals are likely to seek the most specific conclusion consistent with the premises (see Johnson-Laird, 1993), they are likely to seek parsimonious conclusions (see Johnson-Laird & Byrne, 1991), and they are likely to be constrained by the availability of relevant knowledge (Tversky & Kahneman, 1973). The model theory postulates a mechanism for making knowledge progressively available. Reasoners begin by trying to form a model of the current situation, and the retrieval of relevant knowledge is easier if they can form a single model containing all the relevant entities. Once they have formed an initial model, knowledge becomes available to them in a systematic way. They manipulate the spatial or physical aspects of the situation; that is, they manipulate the model directly by procedures corresponding to such changes. Next, they make more abstract conceptual manipulations; for example, they consider the properties of superordinate concepts of entities in the model. Finally, they make still more abstract inferences based on introducing

Mental models and probabilistic thinking

185

relations retrieved from models of analogous situations (cf. Gentner, 1983). Consider the following illustration: Arthur's wallet was stolen from him in the restaurant. The person charged with the offense was outside the restaurant at the time of the robbery. What follows? Reasoners are likely to build an initial model of Arthur inside the restaurant when his wallet is stolen and the suspect outside the restaurant at that time. They will infer that the suspect is innocent. They may then be able to envisage the following sort of sequence of ideas from their knowledge about the kinds of things in the model: (1) Physical and spatial manipulations: The suspect leant through the window to steal the wallet. The suspect stole the wallet as Arthur was entering the restaurant, or ran in and out of the restaurant very quickly (creative inferences that, in fact, are contrary to the premises). (2) Conceptual manipulations: The suspect had an accomplice - a waiter, perhaps - who carried out the crime (theft is a crime, and many crimes are committed by accomplices). (3) Analogical thinking The suspect used a radio-controlled robot to sneak up behind Arthur to take the wallet (by analogy with the use of robots in other "hazardous" tasks). In short, the model theory predicts that reasoners begin by focusing on the initial explicit properties of their model of a situation, and then they attempt to move away from them, first by conceptual operations, and then by introducing analogies from other domains. It is important to emphasize that the order of the three sorts of operations is not inflexible, and that particular problems may elicit a different order of operations. Nevertheless, there should be a general trend in moving away from explicit models to implicit possibilities. Third, reasoners are also likely to be guided by other heuristics, which have been extensively explored by Tversky and Kahneman, and their colleagues. These heuristics can be traced back to Hume's seminal analysis of the connection between ideas: "there appear to be only three principles of connexion between ideas, namely, Resemblance, Contiguity in time or place, and Cause or Effect" (Hume, 1748, Sec. III). Hence, semantic similarity between the premises and the conclusion, and the causal cohesiveness between them, will influence probabilistic judgements. Such factors may even replace extensional estimates based on models.

186

P. Johnson-Laird

Fourth, individuals should be inferential satisficers; that is, if they reach a credible (or desirable) conclusion, or succeed in constructing a model in which such a conclusion is true, they are likely to accept it, and to overlook models that are counterexamples. Conversely, if they reach an incredible (or undesirable) conclusion, they are likely to search harder for a model of the premises in which it is false. This propensity to satisfice will in turn lead them to be overconfident in their conclusions, especially in the case of arguments that do have alternative models in which the conclusion is false. Individuals are indeed often overconfident in their inductive judgements, and Gigerenzer, Hoffrage, and Kleinbolting (1991) have propounded a theory of "probabilistic mental models" to account for this phenomenon. These are long-term representations of probabilistic cues and their validities (represented in the form of conditional probabilities). These authors propose that individuals use the single cue with the strongest validity and do not aggregate multiple cues, and that their confidence derives from the validity of this cue. They report corroboratory evidence from their experiments on the phenomenon of overconfidence; that is, rated confidence tends to be higher than the actual percentage of correct answers. As Griffin and Tversky (1992) point out, however, overconfidence is greater with harder questions and this factor provides an alternative account of Gigerenzer et al.'s results. In contrast, the model theory proposes that the propensity to satisfice should lead subjects to overlook models in the case of multiple-model problems, and so they should tend to be more confident than justified in the case of harder problems. Overconfidence in inductive inference occurred in an unpublished study by Johnson-Laird and Anderson, in which subjects were asked to draw initial conclusions from such premises as: The old man was bitten by a poisonous snake. There was no known antidote available. They tend initially to infer that the old man died. Their confidence in such conclusions was moderately high. They were then asked whether there were any other possibilities and they usually succeeded in thinking of two or three. When they could go no further, they were asked to rate again their initial conclusions, and showed a reliable decline in confidence. Hence, by their own lights, they were initially overconfident, though by the end of the experiment they may have been underconfident as a result of bringing to mind remote scenarios. With easier one-model problems, the error and its correlated overconfidence cannot occur. But should subjects be underconfident in such cases, as is sometimes observed? One factor that may be responsible for the effect in repeated-measure designs is the subjects' uncertainty about whether or not there might be other models in a one-model case. Finally, individuals are likely to focus on what is explicit in their initial models and thus be susceptible to various "focusing effects" (see Legrenzi, Girotto, &

Mental models and probabilistic thinking

187

Johnson-Laird, 1993). These effects include difficulty in isolating genuinely diagnostic data (see, for example, Beyth-Marom & Fischhoff, 1983; Doherty, Mynatt, Tweney, & Schiavo, 1979), testing hypotheses in terms of their positive instances (Evans, 1989; Klayman & Ha, 1987), neglect of base rates in certain circumstances (Tversky & Kahneman, 1982), and effects of how problems in deductive and inductive reasoning are framed (e.g., Johnson-Laird & Byrne, 1989; Tversky & Kahneman, 1981). Focusing is also likely to lead to too great a reliance on the credibility of premises (and conclusion) and too little on the strength of the argument, that is, the relation between the premises and conclusion. Reasoners will build an initial model that makes explicit the case for a conclusion, and then fail to adjust their estimates of its likelihood by taking into account alternative models (see also Griffin & Tversky, 1992, for an analogous view). Conversely, any factor that makes it easier for individuals to flesh out explicit models of the premises should improve performance.

6. Rules for probabilistic thinking An obvious potential basis for probabilistic reasoning is the use of rules of inference, such as: If q & r then s (with probability p) .'. If q then s (with probability p') Numerous AI programs include rules of this sort (see, for example, Holland, Holyoak, Nisbett, & Thagard, 1986; Michalski, 1983; Winston, 1975). The most plausible psychological version of this idea is due to Collins and Michalski (1989). They argue that individuals construct mental models on the basis of rules of inference, and that these rules have numerical parameters for such matters as degree of certainty. They have not tried to formalize all patterns of plausible inference, but rather some patterns of inference that make up a core system of deductions, analogies and inductions. They admit that it is difficult to use standard psychological techniques to test their theory, which is intended to account only for people's answers to questions. It does not make any predictions about the differences in difficulty between various sorts of inference, and, as they point out (p. 7), it does not address the issue of whether people make systematic errors. Hence, their main proposed test consists in trying to match protocols of arguments against the proposed forms of rules. Pennington and Hastie (1993) report success in matching these patterns to informal inferences of subjects playing the part of trial jurors. But, as Collins and Michalski mention, one danger is that subjects' protocols are merely rationalizations for answers arrived at by other means. In sum, AI rule systems for induction have not yet received decisive corroboration.

188

P. Johnson-Laird

In contrast, another sort of rule theory has much more empirical support. This theory appeals to the idea that individuals have a tacit knowledge of such rules as the "law of large numbers" (see Nisbett, 1993; Smith, Langston, & Nisbett, 1992). Individuals apply the rules to novel materials, mention them in justifying their responses, benefit from training with them, and sometimes overextend their use of them. The rules in AI programs are formal and can be applied to the representation of the abstract logical form of premises. The law of large numbers, however, is not a formal rule of inference. It can be paraphrased as follows: The larger the sample from a population the smaller its mean is likely to diverge from the population mean. Aristotle would not have grasped such notions as sample, mean and population, but he would have been more surprised by a coin coming up heads ten times in a row than a coin coming up heads three times in a row. He would thus have had a tacit grasp of the law that he could make use of in certain circumstances. The law has a rich semantic content that goes well beyond the language of logical constants, and it is doubtful whether it could be applied to the logical form of premises. On the contrary, it is likely to be applied only when one has grasped the content of a problem, that is, constructed a model that makes explicit that it calls for an estimate based on an example. Individuals are likely to hold many other general principles as part of their beliefs about probability. For instance, certain devices produce different outcomes on the basis of chance, that is, at approximately equal rates and in unpredictable ways; if a sample from such a device is deviant, things are likely to even up in the long run (gambler's fallacy). Such principles differ in generality and validity, but they underlie the construction of many probabilistic judgements. The fact that individuals can be taught correct laws and that they sometimes err in over-extending them tells us nothing about the mental format of the laws. They may take the form of schemas or content-specific rules of inference, but they could be represented declaratively. Likewise, how they enter into the process of thinking - the details of the computations themselves - is also unknown. There is, however, no reason to oppose them to mental models. They seem likely to work together in tandem, just as conceptual knowledge must underlie the construction of models.

7. Conclusions The principle thesis of the present paper is that general knowledge and beliefs, along with descriptions of situations, lead to mental models that are used to assess probabilities. Most cognitive scientists agree that humans construct mental

Mental models and probabilistic thinking

189

representations; many may suspect that the model theory merely uses the words "mental model" where "mental representation" would do. So, what force, if any, is there to the claim that individuals think probabilistically by manipulating models? The answer, which has been outlined here, is twofold. First, the representational principles of models allow sets of possibilities to be considered in a highly compressed way, and even in certain cases sets of sets of possibilities. Hence, it is feasible to assess probability by estimating possible states of affairs within a general framework that embraces deduction, induction and probabilistic thinking. This framework provides an extensional foundation of probability theory that is not committed a priori to either a frequency or degrees-of-belief interpretation, which are both equally feasible on this foundation. Second, the model theory makes a number of predictions based on the distinction between explicit and implicit information, and on the processing limitations of working memory. Such predictions, as the study of deduction has shown, are distinct from those made by theories that postulate only representations of the logical form of assertions.

References Aristotle (1984). The complete works of Aristotle, edited by J. Barnes, 2 vols. Princeton: Princeton University Press. Bar-Hillel, Y., & Carnap, R. (1964). An outline of a theory of semantic information. In Y. Bar-Hillel, (Ed.), Language and information. Reading, MA: Addison-Wesley. Barwise, J. (1993). Everyday reasoning and logical inference. Behavioral and Brain Sciences, 16, 337-338. Barwise, J., & Etchemendy, J. (1989). Model-theoretic semantics. In M.I. Posner (Ed.), Foundations of cognitive science. Cambridge, MA: MIT Press. Beyth-Marom, R., & Fischhoff, B. (1983). Diagnosticity and pseudodiagnosticity. Journal of Personality and Social Psychology, 45, 1185-1197. Braine, M.D.S., Reiser, B.J., & Rumain, B. (1984). Some empirical justification for a theory of natural propositional logic. In The psychology of learning and motivation, (Vol. 18). New York: Academic Press. Byrne, R.M.J. (1989). Suppressing valid inferences with conditionals. Cognition, 31, 61-83. Carnap, R. (1950). Logical foundations of probability. Chicago: Chicago University Press. Collins, A.M., & Michalski, R. (1989). The logic of plausible reasoning: A core theory. Cognitive Science, 13 1-49. Craik, K. (1943). The nature of explanation. Cambridge, UK: Cambridge University Press. Doherty, M.E., Mynatt, C.R., Tweney, R.D., & Schiavo, M.D. (1979). Pseudodiagnosticity. Acta Psychologica, 43, 11-21. Evans, J.St.B.T. (1989). Bias in human reasoning: Causes and consequences. London: Erlbaum. Fisher, A. (1988). The logic of real arguments. Cambridge, UK: Cambridge University Press. Garnham, A. (1987). Mental models as representations of discourse and text. Chichester: Ellis Horwood. Garnham, A. (1993). A number of questions about a question of number. Behavioral and Brain Sciences, 16, 350-351. Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155-170.

190

P. Johnson-Laird

Gigerenzer, G., Hoffrage, U., & Kleinbolting, H. (1991). Probabilistic mental models: A Brunswikian theory of confidence. Psychological Review, 98, 506-528. Griffin, D., & Tversky, A. (1992). The weighing of evidence and the determinants of confidence. Cognitive Psychology, 24, 411-435. Hempel, C. (1965). Aspects of scientific explanation. New York: Macmillan. Hintikka, J. (1962). Knowledge and belief: An introduction to the logic of the two notions. Ithaca: Cornell University Press. Hintzman, D.L., Nozawa, G., & Irmscher, M. (1982). Frequency as a nonpropositional attribute of memory. Journal of Verbal Learning and Verbal Behavior, 21, 127-141. Holland, J.H., Holyoak, K.J., Nisbett, R.E., & Thagard, P. (1986). Induction: Processes of inference, learning, and discovery. Cambridge, MA: MIT Press. Hume, D. (1748/1988). An enquiry concerning human understanding. La Salle, IL: Open Court. Johnson-Laird, P.N. (1983). Mental models: Towards a cognitive science of language, inference and consciousness. Cambridge, UK: Cambridge University Press. Johnson-Laird, P.N. (1993). Human and machine thinking. Hillsdale, NJ: Erlbaum. Jbhnson-Laird, P.N., & Bara, B. (1984). Syllogistic inference. Cognition, 16, 1-61. Johnson-Laird, P.N., & Byrne, R.M.J. (1989). Only reasoning. Journal of Memory and Language, 28, 313-330. Johnson-Laird, P.N., & Byrne, R.M.J. (1991). Deduction. Hillsdale, NJ: Erlbaum. Johnson-Laird, P.N., & Byrne, R.M.J. (1993). Authors' response [to multiple commentaries on Deduction]: Mental models or formal rules? Behavioral and Brain Sciences, 16, 368-376. Kahneman, D., & Tversky, A. (1972). Subjective probability: A judgment of representativeness. Cognitive Psychology, 3, 430-454. Klayman, J., & Ha, Y.-W. (1987). Confirmation, disconfirmation and information in hypothesis testing. Psychological Review, 94, 211-228. Legrenzi, P., Girotto, V, & Johnson-Laird, P.N. (1993). Focussing in reasoning and decision making. Cognition, 49, 37-66. Marr, D. (1982). Vision: A computational investigation into the human representation and processing of visual information. San Francisco: W.H. Freeman. Medvedev, Z.A. (1990). The legacy of Chernobyl. New York: Norton. Michalski, R.S. (1983). A theory and methodology of inductive learning. In R.S. Michalski, J.G. Carbonell, & T.M. Mitchell (Eds.), Machine learning: An artificial intelligence approach. Los Altos, CA: Morgan Kaufmann. Montague, R. (1974). Formal philosophy: Selected papers. New Haven: Yale University Press. Nisbett, R.E. (Ed.) (1993). Rules for reasoning. Hillsdale, NJ: Erlbaum. Oakhill, J.V., Johnson-Laird, P.N. & Garnham, A. (1989). Beiievability and syllogistic reasoning. Cognition, 31, 117-140. Osherson, D.N., Smith, E.E. & Shafir, E. (1986). Some origins of belief. Cognition, 24, 197-224. Pennington, N., & Hastie, R. (1993). Reasoning in explanation-based decision making. Cognition, 49, 123-163. Polk, T.A., & Newell, A. (1988). Modeling human syllogistic reasoning in Soar. In Tenth Annual Conference of the Cognitive Science Society (pp. 181-187). Hillsdale, NJ: Erlbaum. Ramsay, F.P. (1926/1990). Truth and probability. In D.H. Mellor, (Ed.), F.P. Ramsay: Philosophical papers. Cambridge, UK: Cambridge University Press. Rips, L.J. (1983). Cognitive processes in propositional reasoning. Psychological Review, 90, 38-71. Shafer, G. (1993). Using probability to understand causality. Unpublished MS, Rutgers University. Shafer G., & Tversky, A. (1985). Languages and designs for probability judgment. Cognitive Science, 9, 309-339. Shafir, E., & Tversky, A. (1992). Thinking through uncertainty: Nonconsequential reasoning and choice. Cognitive Psychology, 24, 449-474. Smith, E.E., Langston, C, & Nisbett, R.E. (1992). The case for rules in reasoning. Cognitive Science, 16, 1-40. Thagard, P. (1989). Explanatory coherence. Behavioral and Brain Sciences, 12, 435-502. Toulmin, S.E. (1958). The uses of argument. Cambridge, UK: Cambridge University Press.

Mental models and probabilistic thinking

191

Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5, 207-232. Tversky, A., & Kahneman, D. (1981). The framing of decisions and the psychology of choice. Science, 211, 453-458. Tversky, A., & Kahneman, D. (1982). Evidential impact of base rates. In D. Kahneman, P. Slovic, & A. Tversky, (Eds.), Judgments under uncertainty: Heuristics and biases. New York: Cambridge University Press. Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment. Psychological Review, 90, 293-315. Wason, P.C. (1965). The contexts of plausible denial. Journal of Verbal Learning and Verbal Behavior, 4, 7-11. Winston, P.H. (1975). Learning structural descriptions from examples. In P.H. Winston, (Ed.), The psychology of computer vision. New York: McGraw-Hill.

10 Pretending and believing: issues in the theory of ToMM Alan M. Leslie Department of Psychology, Center for Cognitive Science, Rutgers University, Piscataway, NJ 08855-1179, USA

Abstract Commonsense notions of psychological causality emerge early and spontaneously in the child. What implications does this have for our understanding of the mind/brain and its development? In the light of available evidence, the child's "theory of mind" is plausibly the result of the growth and functioning of a specialized mechanism (ToMM) that produces domain-specific learning. The failure of early spontaneous development of "theory of mind" in childhood autism can be understood in terms of an impairment in the growth and functioning of this mechanism. ToMM constructs agent-centered descriptions of situations or "metarepresentations". Agent-centered descriptions place agents in relation to information. By relating behavior to the attitudes agents take to the truth of propositions, ToMM makes possible a commonsense causal interpretation of agents9 behavior as the result of circumstances that are imaginary rather than physical. Two early attitude concepts, pretends and believes, are discussed in the light of some current findings. Dedication: This article is dedicated to the memory of Daniel Roth, my student, collaborator and friend who tragically lost his long struggle against cancer on April 17, 1993.

This paper has undergone a long gestation, various parts having been presented to the BPS Developmental Section Annual Conference, Coleg Harlech, September 1988, Trieste Encounters on Cognitive Science, Trieste, Italy, May 1989, International Workshop on Naturalized Epistemology, Cornell University, December 1989, International Conference on Cultural Knowledge and Domain Specificity, University of Michigan, Ann Arbor, October 1990, Society for Research on Child Development Biennial Meeting, Seattle, April 1991, and Inaugural Conference of the Rutgers University Center for Cognitive Science, Rutgers University, November 1991. I am grateful to participants and audiences at those meetings and also to colleagues and friends at the MRC Cognitive Development Unit for nurture and good nature.

194

A. Leslie

oAnn

A \

A box


p])> -5. We would like to represent the fit-relation in terms of vectors. For this purpose, suppose that n-dimensional vectors are assigned to obj Upred, one per object and one per predicate. Given such an assignment, let us say that o "dominates" p just in case the coordinates of o's vector are at least as great as the corresponding coordinates of /?'s vector. We have the following fact, proved in Doignon, Ducamp, and Falmagne (1984): (4) Let P embody any fit-relation whatsoever. Then for some n, there is an assignment of n-dimensional vectors to obj U pred such that for all o E obj and all p E pred, P([o, p]) > .5 if and only if o dominates p. Moreover, n can be chosen to not exceed the smaller of: the cardinality of obj, the cardinality of pred. Intuitively, we can think of the vector assigned to a predicate as establishing criteria for membership in the associated category. (For example, the predicate "can learn a four choice-point maze in three trials" might have a requirement of .75 in the coordinate corresponding to intelligent.) For o to have greater than .5 probability of possessing p, oJs values at each coordinate must exceed the criterion established by p. Fact (4) shows that such a scheme is perfectly general for representing probability thresholds and it renders plausible the idea that real vectors might also serve to predict continuous assessments of the probability of statements.

3.3. To and from statement representations Recall that our goal is to capture the coherent core of a person's judgement about chances. Call the person at issue 5if. Having decided to use vectors to represent obj U pred, two questions remain to be answered. These are: (a) Which particular object and predicate are attributed to 5if? (b) How are object and predicate vectors translated into a probability distribution? Once answers are offered to (a) and (b), a third question may be addressed, namely: (c) If 2fs vectors are fixed in accordance with the answer to (a), and if

D. Osherson, E. Shafir, E. Smith

230

probabilities are subsequently assigned to propositions in accordance with the answer to (b), how well do the resulting probabilities accord with 2T$ judgement about chances? Does the processed and regimented output of our system retain any of the insight that characterizes 2Ts understanding about probabilities in the environment? Let us now briefly consider (a)-(c).

3.4. Fixing object and predicate vectors One means of obtaining vectors is to request the needed information directly from %t via feature ratings (as in Osherson, Stern, Wilkie, Stob, & Smith, 1991). A less direct approach is to infer the vectors from similarity ratings among objects and predicates. In this case, we work backwards from a suitable vector-based model of similarity (e.g., those discussed in Osherson, 1987; Suppes, Krantz, Luce, & Tversky, 1989; Tversky, 1977), looking for vectors that best predict 2Ts similarity data. Another strategy is to postulate a model of simple probability judgements based on the needed vectorial representations, and then work backwards to vectors from such judgements. In this case, our system carries out "extrapolation", extending a small set of probability judgements to a more complete set (see Osherson, Smith, Meyers, Shafir, & Stob, in press).

3.5. Vectors to probabilities Turning to question (b) above, we describe one procedure for synthesizing probabilities from the vectors underlying obj and pred. It rests upon a scheme for constructing three-dimensional Venn diagrams. Specifically, the pair of vectors associated with object o and predicate p is translated into a subregion 5? of the unit cube.3 The volume of 9t represents the probability of [o, p]. The position of 91 determines its intersection with subregions assigned to other statements. The probability of a complex proposition (e.g., the intersection or the union of two statements) may then be determined by calculating the volume of the corresponding region. It is easy to see that use of the diagram guarantees probabilistic coherence.4 Let us now outline a simple scheme for selecting the particular region assigned 3 The unit cube has sides of length 1. It is used for convenience in what follows; various other kinds of solids would serve as well. 4 There is no mathematical reason to limit the diagram to three dimensions. Volumes in the n-dimensional unit cube for any positive n yield bona fide distributions. So far our experiments indicate that three dimensions are enough.

Extracting the coherent core of human probability judgment

231

to a given statement [o, p]. Let O, P be the vectors underlying o, /?, respectively, and suppose them to be suitably normalized so that all coordinates fall into the interval [0,1]. Define the "O, P-box" to be the (unique) rectangular solid 9? with the following properties: (5) (a) O falls within 38; (b) for i ^ 3 the length of the ith side of 5? is 1 minus the absolute difference between Oi and P,; (c) within the foregoing constraints, P is as close as possible to the geometrical center of 01. It may be seen that the volume of 9t varies directly with the match of P's coordinates to those of 0\ statements based on compatible objects and predicates are accorded higher probability thereby. Moreover, <Jfc's position in the cube represents aspects of the semantics of [o9 p]. For example, if p and q are complementary predicates with contrasting vectors then (5) assigns [o, p] and [o, q] boxes with little or no intersection. This reflects the low probability that must sensibly be assigned to [o, p] A [O, q] in view of the incompatible contents of p and q. Many alternatives to (5) are possible. To serve as a computationally tractable means of coherent reasoning in large domains it suffices to meet the following condition: (C) Given a point x in the unit cube, and vectors 0, P underlying object o and predicate /?, it must be computationally easy to determine whether x lies in the region associated with [o, p]. In this case it is straightforward to calculate the volumes associated with any Boolean combination of statements, hence with any proposition.5 It is clear that (5) satisfies C. Observe that within any Venn diagram scheme that conforms to C, coherent probabilistic reasoning can proceed without storing an entire distribution. It is enough to store the vectors underlying objects and predicates since the volume associated with any given proposition (of reasonable length) can be easily retrieved from the vectors. Thus, given 10 objects 2nd 10 predicates, only 20 vectors need be stored. This is easily achieved even for vectors of considerable size. In contrast, 10 objects and 10 predicates give rise to 100 statements and thus to a distribution with 2100 state descriptions. A potential solution to the problem of coherent reasoning, posed in section 2.4, is offered thereby. It must be emphasized that not every distribution can be represented by a Venn A more careful formulation of C would refer to € -spheres in place of points x, etc.

232

D. Osherson, E. Shafir, E. Smith

diagram that meets C (just as not every distribution manifests conditional independencies of a computationally convenient kind). The question thus arises: do distributions that conform to C approximate human intuition about chance in a wide variety of domains? We are thus led to question (c) above, namely, whether the distribution delivered by our method resembles the original intuitions of subject $?. Sticking with the simple scheme in (5) - henceforth called the "Venn model" - let us now address this matter.

3.6. Accuracy of the method We summarize one experimental test of our method. By an elementary argument (over obj U pred) is meant a non-empty set of statements, one of which is designated as "conclusion", the remainder (if any) as "premises". Statements are considered special cases of elementary arguments, in which the premise set is empty. An argument may be conceived as an invitation to evaluate the probability of its conclusion while assuming the truth of its premises. Thirty college students evaluated 80 elementary arguments based on four mammals (which served as objects) and two predicates (e.g., "are more likely to exhibit 'fight' than 'flight' posture when startled"). For each subject, an individually randomized selection of 30 arguments was used to fix vectors representing his objects and predicates. This was achieved by working backwards from the Venn model, seeking vectors that maximize its fit to the subject's judgement about the 30 input arguments. The Venn model was then applied to the resulting vectors to produce probabilities for the remaining 50 arguments. For each subject we calculated the average, absolute deviation between the Venn model's predictions for the 50 arguments and the probabilities offered directly by the subject. Pearson correlations between the two sets of numbers were also calculated. The median, average absolute deviation between the observed probabilities assigned to a subject's 50 predicted arguments and the probabilities generated by the Venn model is .ll. 6 The correlation between the two sets of numbers is .78. The results suggest that the Venn method can extrapolate a coherent set of probabilities from a small input set, and do this in such a way that the extrapolated distribution provides a reasonable approximation to the judgement of the person providing input. The input set of probabilities need not be coherent.

6 This deviation can be compared to the following statistic. Consider the mean of the probabilities assigned to the 30 arguments used to fix the object and predicate vectors of a given subject. We may use this single number as a predictor of the probabilities assigned to the remaining 50 arguments. In this case the median, average absolute deviation between the observed and predicted probabilities is .20.

Extracting the coherent core of human probability judgment

233

4. Concluding remarks The problem of recovering the coherent core of human probability judgement strikes us as an important project for cognitive psychology. It unites theorizing about the mental mechanisms of reasoning with a practical problem for expert systems, namely, finding an exploitable source of Bayesian priors. The system sketched above is preliminary in character, and serves merely to suggest the feasibility of the research program we advocate. Psychological research in recent years has produced considerable understanding of the character and causes of incoherent reasoning, even if debate continues about its scope and interpretation (see Gigerenzer & Murray, 1987; Osherson, 1990; Shafir & Tversky, 1992; Tversky & Shafir, 1992, and references cited there). It was noted in section 2.1 that probabilistic coherence has non-trivial justification as a standard - however incomplete - of normatively acceptable reasoning. We thus take there to be good empirical evidence, plus great computational plausibility, in favor of the thesis that human judgement is imperfect from the normative point of view. This thesis does not, however, impugn every aspect of ordinary reasoning. Indeed, the merits of human judgement have often been emphasized by the very researchers who investigate its drawbacks (e.g., Nisbett & Ross, 1980, p. 14). A challenge is posed thereby, namely, to devise methods that distill the rational component of human thought, isolating it from the faulty intuition that sometimes clouds our reason. Such appears to have been the goal of early inquiry into probability and utility (Gigerenzer et al., 1989, Ch. 1). It remains a worthy aim today.

References Andreassen, S., Woldbye, M., Falck, B., & Andersen, S. (1989). Munin: A causal probabilistic network for interpretation of electromyographic findings. In Proceedings of the Tenth International Joint Conference on Artificial Intelligence. Bobrow, D., & Winograd, T. (1976). An overview of KRL, a knowledge representation language. Cognitive Science, 1, 3-46. Casscells, W., Schoenberger, A., & Grayboys, T. (1978). Interpretation by physicians of clinical laboratory results. New England Journal of Medicine, 299, 999-1000. Cooper, G.F. (1987). Probabilistic inference using belief networks is np-hard. Memo KSL-87-27, Knowledge Systems Laboratory, Stanford University, May 1987. Cox, R. (1946). Probability, frequency, and reasonable expectation. American Journal of Physics, 14, 1-13. de Finetti, B. (1964). La prevision: Ses lois logiques, ses sources subjectives [transl. into English]. In H. Kyburg & P. Smokier (Eds.), Studies in subjective probability. New York: Wiley. de Finetti, B. (1972). Probability, induction and statistics. New York: Wiley. Doignon, J.-P., Ducamp, A., & Falmagne, J.-C. (1984). On realizable biorders and the biorder dimension of a relation. Journal of Mathematical Psychology, 28, 73-109. Dubois, D., & Prade, H. (1991) Updating with belief functions, ordinal conditional functions and

234

D. Osherson, E. Shafir, E. Smith

possibility measures. In P.P. Bonissone, M. Henrion, L.N. Kanal, & J.F. Lemmer (Eds.), Proceedings of the Sixth Workshop on Uncertainty in Artificial Intelligence (pp. 311-329). Amsterdam: Elsevier. Fagin, R., & Halpern, J. (1991). A new approach to updating beliefs. In P.P. Bonissone, M. Henrion, L.N. Kanal, & J.F. Lemmer (Eds.), Proceedings of the Sixth Workshop on Uncertainty in Artificial Intelligence (pp. 347-374). Amsterdam: Elsevier. Gaifman, H. & Snir, M. (1982). Probabilities over rich languages. Journal of Symbolic Logic, 47, 495-548. Geiger, D., Verma, T., & Pearl, J. (1990). d-Separation: From theorems to algorithms. In R.D. Shachter, M. Henrion, L.N. Kanal, & J.F. Lemmer (Eds.), Uncertainty in artificial intelligence 5. Amsterdam: North-Holland. Gigerenzer, G., & Murray, D.J. (1987). Cognition as intuitive statistics. Hillsdale, NJ: Erlbaum. Gigerenzer, G., Swijtink, Z., Porter, T., Daston, L., Beatty, J., & Kruger, L. (1989). The empire of chance. Cambridge, UK: Cambridge University Press. Glymour, C. (1992). Thinking things through. Cambridge, MA: MIT Press. Heckerman, D. (1986). Probabilistic interpretations for MYCIN'S certainty factors. In L.N. Kanal & J.F. Lemmer (Eds.), Uncertainty in artificial intelligence. Amsterdam: North-Holland. Heckerman, D. (1990a). Probabilistic similarity networks. Heckerman, D. (1990b). A tractable inference algorithm for diagnosing multiple diseases. In R.D. Schachter, M. Henrion, L.N. Kanal, & J.F. Lemmer (Eds.), Uncertainty in artificial intelligence 5. Amsterdam: North-Holland. Jeffrey, R. (1983). The logic of decision (2nd edn.). Chicago: University of Chicago Press. Kahneman, D., Slovic, P., & Tversky, A. (Eds.) (1982). Judgment under uncertainty: Heuristics and biases. Cambridge, UK: Cambridge University Press. Kahneman, D., & Tversky, A. (1972). Subjective probability: A judgement of representativeness. Cognitive Psychology, 3, 430-454. Kelly, K.T. (1993). The logic of reliable inquiry. MS, Department of Philosophy, Carnegie Mellon University. Lauritzen, S., & Spiegelhalter, D. (1988). Local computations with probabilities on graphical structures and their applications to expert systems. Journal of the Royal Statistical Society, B, 50, 157-224. Lindley, D. (1982). Scoring rules and the inevitability of probability. International Statistical Review, 50, 1-26. Long, W., Naimi, S., Criscitiello, M., & Jayes, R. (1987). The development and use of a causal model for reasoning about heart failure. IEEE Symposium on Computer Applications in Medical Care (pp. 30-36). Minsky, M. (1981). A framework for representing knowledge. In J. Haugeland (Ed.), Mind design. Cambridge, MA: MIT Press. Minsky, M. (1986). The society of mind. New York: Simon & Schuster. Neapolitan, R. (1990). Probabilistic reasoning in expert systems. New York: Wiley. Nisbett, R., & Ross, L. (1980). Human inference: Strategies and shortcomings of social judgment. Englewood Cliffs, NJ: Prentice-Hall. Olesen, K.G., Kjaerulff, U., Jensen, F., Jensen, F.V., Falck, B., Andreassen, S., & Andersen, S.K. (1989). A munin network for the median nerve: A case study on loops. Applied Artificial Intelligence, Special Issue: Towards Causal AI Models in Practice. Osherson, D. (1987). New axioms for the contrast model of similarity. Journal of Mathematical Psychology, 31, 93-103. Osherson, D. (1990). Judgment. In D. Osherson & E.E. Smith (Eds.), Invitation to cognitive science: Thinking (pp. 55-88). Cambridge, MA: MIT Press. Osherson, D. (in press). Probability judgment. In E.E. Smith, & D. Osherson (Eds.), Invitation to cognitive science: Thinking (2nd edn.). Cambridge, MA: MIT Press. Osherson, D., Biolsi, K., Smith, E., Shafir, E., & Gualtierotti, A. (1994). A source of Bayesian priors. IDIAP Technical Report 94-03. Osherson, D., Smith, E.E., Meyers, T.S., Shafir, E., & Stob, M. (in press). Extrapolating human probability judgment. Theory and Decision.

Extracting the coherent core of human probability judgment

235

Osherson, D., Stern, J., Wilkie, O., Stob, M., & Smith, E. (1991). Default probability, Cognitive Science, 15, 251-270. Pearl, J. (1986). Fusion, propogation, and structuring in belief networks. Artificial Intelligence, 29, 241-288. Pearl, J. (1988). Probabilistic reasoning in intelligent systems: Networks of plausible inference. San Mateo, CA: Morgan-Kaufmann. Resnik, M. (1987). Choices: An introduction to decision theory. Minneapolis: University of Minnesota Press. Ross S. (1988). A first course in probability. New York: Macmillan. Savage L. (1972). The foundations of statistics. New York: Dover. Shafter, G. (1976). A mathematical theory of evidence. Princeton, NJ: Princeton University Press. Shafer, G. (1986). Constructive probability. Synthese, 48, 1-60. Shafer, G., & Pearl, J. (Eds.) (1990). Readings in uncertain reasoning. San Mateo, CA: MorganKaufmann. Shafir, E., & Tversky, A. (1992). Thinking through uncertainty: Nonconsequential reasoning and choice. Cognitive Psychology, 24, 449-414. Shortliffe, E., & Buchanan, B. (1975). A model of inexact reasoning in medicine. Mathematical Biosciences, 23, 351-379. Smith, E.E., & Medin, D. (1981). Categories and concepts. Cambridge, MA: Harvard University Press. Stockmeyer, L. (1974). The complexity of decision problems in automata theory and logic. Ph.D. thesis, MIT. Suppes, P., Krantz, D.H., Luce, R.D., & Tversky, A. (1989). Foundations of measurement (Vol. II). San Diego: Academic Press. Szolovits, P., & Pauker, S. (1978). Categorical and probabilistic reasoning in medical diagnosis. Artificial Intelligence, 11, 115-144. Tversky, A. (1977). Features of similarity. Psychological Review, 84, 327-362. Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The conjunction fallacy in probability judgement. Psychological Review, 90, 293-315. Tversky, A., & Shafir, E. (1992). The disjunction effect in choice under uncertainty. Psychological Science, 5, 305-309. Whittaker, J. (1990). Graphical models in applied multivariate statistics. New York: Wiley. Winterfeld, D.V., & Edwards, W. (1986). Decision analysis and behavioral research. New York: Cambridge University Press.

12 Levels of causal understanding in chimpanzees and children David Premack*, Ann James Premack Laboratoire de Psycho-Biologie du Developpement, CNRS, 41 rue Gay Lussac, 75005 Paris, France

Abstract We compare three levels of causal understanding in chimpanzees and children: (1) causal reasoning, (2) labelling the components (actor, object, and instrument) of a causal sequence, and (3) choosing the correct alternative for an incomplete representation of a causal sequence. We present two tests of causal reasoning, the first requiring chimpanzees to read and use as evidence the emotional state of a conspecific. Despite registering the emotion, they failed to use it as evidence. The second test, comparing children and chimpanzees, required them to infer the location of food eaten by a trainer. Children and, to a lesser extent, chimpanzees succeeded. When given information showing the inference to be unsound - physically impossible - 4-year-old children abandoned the inference but younger children and chimpanzees did not. Children and chimpanzees are both capable of labelling causal sequences and completing incomplete representations of them. The chimpanzee Sarah labelled the components of a causal sequence, and completed incomplete representations of actions involving multiple transformations. We conclude the article with a general discussion of the concept of cause, suggesting that the concept evolved far earlier in the psychological domain than in the physical.

•Corresponding author. The data reported here were collected at the University of California, Santa Barbara, and the University of Pennsylvania Primate Facility. We are greatly indebted to Guy Woodruff, who participated in all phases of the research and would be a co-author if we knew his whereabouts and could obtain his permission. We are also indebted to the many students, graduate and undergraduate, at both institutions who assisted in the care and testing of the chimpanzees.

238

D. Premack, A. Premack

Introduction In this paper we compare three levels of causal understanding in chimpanzees and children. At the deepest level, the individual engages in causal reasoning, solving problems in which he sees the outcome of a process but not its cause, and must infer or reconstruct the missing events. At an intermediate level the individual analyses intact causal sequences into their components and labels them. He must label the actor, object, and instrument of the action. The ability to carry out this task is a prerequisite for causal reasoning; if one cannot identify the separate parts of an intact sequence, one cannot identify the missing part of an incomplete sequence. At the most superficial level, the individual must complete an incomplete representation of a causal action, by selecting the correct alternative. Since the alternatives are all visible this task is the least demanding. Chimpanzees have been shown to be capable of all but the deepest level (Premack 1976, 1983), though they have been shown capable of analogical reasoning. Not only do they complete incomplete analogies and make same/ different judgments about exemplars that are and are not analogies (Gillan, Premack, & Woodruff, 1981), they also construct analogies from scratch (Oden & Premack, unpublished data). There is evidence (Gillan, 1981), albeit inconclusive, that they can do transitive inference. But there is little indication that they are capable of "Sherlock Holmes" type reasoning, that is, causal reasoning. In causal reasoning, an outcome-a corpse on the floor-is presented, but the cause is not. A human confronted with this scene would ask immediately, "Who did it? How? When, where and why?" He would answer these questions by making inferences from the "evidence", which has two main sources: existing knowledge concerning the "corpse" and its past, and observations concerning the immediate circumstances. As an astute observer, Sherlock Holmes was a good reasoner because he had an uncanny sense of what was "relevant", detecting implications in what others dismissed as neutral facts.

Causal reasoning In this paper we present two tests of causal reasoning: one conducted with a group of chimpanzees, another in which we compare chimpanzees and young children.

Levels of causal understanding in chimpanzees and children

239

Subjects The chimpanzees (Pan troglodytes) were African born; four were 3^-4\ years old and Sarah was about 10 years old at the time of the study. Animals entered the laboratory as infants, were diapered and bottle-fed, trained extensively in match-to-sample, and (some) were taught an artificial language. They were maintained rather like middle-class children on three meals a day, snacks, and continuous access to water. When tested, some were "rewarded" with a preferred food while others were simply praised. Participating children came from Philadelphia schools, and varied in age from 3.8 to 4.5 years, with an average age of 4.1 years.

Reading emotional evidence Four juvenile chimpanzees were tested in a simple two-step reasoning problem. We first trained them to run down a path to a consistently positive hidden object, then introduced them to occasional negative trials; that is, a rubber snake was substituted for the food on 15% of trials on a random schedule. The unpredictable negative trials profoundly changed the character of the chimpanzees' run. They no longer dashed full speed to the goal, but slowed midway, approaching the goal hesitantly. We next offered the animals an opportunity to play Holmes, to escape the uncertainty of the negative trial by using the emotional state of an informant to infer what object it had encountered on its run. Now, before starting a run, each animal was placed in a holding room with an informant that had just completed a run. The informant, having encountered either food or the rubber snake on its run, was in either a positive or negative emotional state, and this state, we have reason to believe, was successfully communicated to the recipient. Uninformed human judges shown videotapes of the informant could discriminate the state of the informant (ca. 98% correct). Beyond that, they could discriminate the recipient's state following its contact with the informant (ca. 70% correct). However, the "informed" chimpanzees seemed not to profit from this contact, for they accepted all opportunities to run, and did so in the same way whether: (1) the informant was in a positive state, (2) a negative state, or even when on control trials (3) had had no contact with an informant at all. The holding room was adjacent to the runway. Animals were taken there prior to, and immediately after, a run (to serve as an informant). Every chimpanzee played both roles, that of informant and recipient, and had the opportunity to observe that its conspecifics too played both roles. The use of four animals permitted 12 possible recipient-informant pairs, all of which were used.

240

D. Premack, A. Premack

Under these conditions, the animals should have been able to infer that an informant's emotional state was the result of what it had found at the end of a run. No chimpanzee ever encountered another in the holding room whose emotional state was not owed to a run. Nonetheless, it still could not use the emotional state (which it registered at some level) as evidence from which to infer what the informant had encountered on its run. Could the chimpanzee have made this inference but not have assumed that what the informant found was not a good prediction of what it would find - in other words, assumed it might be snakes for you but food for me? We cannot rule out this possibility, but a human in this circumstance would certainly explore the hypothesis that snakes for you means snakes for me as well. Before testing the chimpanzees, we had assumed this was a simple problem, one that could be easily solved, and therefore could be used as a base condition on which to impose variations that would permit our answering fundamental questions about reasoning. Perhaps at 3±-4± years the chimpanzees were too young and could have solved the problem when older. The age at which children can solve this problem is not known, for one cannot test children with frightening objects.

Using location as evidence In the next experiment, we tested both chimpanzees (the same group of four used in the previous problem), and two groups of children (10 in each group). The apes were tested in their outdoor compound, and the children in a classroom. For the chimpanzees, we placed two opaque containers about 30 feet apart. These formed the base of a triangle with the chimpanzee at its apex, midway between the containers and 30 feet from the base. The chimpanzee was accompanied by a trainer. As the two watched, a second trainer placed an apple in one container and a banana in the other. Following this action, the accompanying trainer placed a blind in front of the chimpanzee so that the containers were no longer visible to it. The trainer distracted the animal for approximately 2 min before removing the blind. What the subject now saw was the second trainer standing midway between the containers eating either an apple or a banana. Having eaten the fruit, the trainer left, and the chimpanzee was released. Each animal was given 10 trials, with an intertrial interval of about an hour. The two fruit were placed equally often in both containers, and the fruit eaten by the experimenter was counterbalanced over trials. We used apples and bananas because apes are fond of both and find them about equally desirable. Children were tested with a comparable procedure adjusted to suit a classroom and human food preferences.

Levels of causal understanding in chimpanzees and children

241

Results Children in both groups were largely consistent, 18 of 20 choosing the container associated with an item different from the one the experimenter was eating.1 That is, on trials on which the experimenter was seen eating a cookie, children selected the container which held the doughnut, and on trials on which he was seen eating the doughnut, selected the container which held the cookie. Chimpanzees, however, were inconsistent. Sadie, the oldest chimpanzee, responded as did the children, choosing the container associated with a fruit different from the one eaten by the experimenter. She made this selection on the first trial as well as on all subsequent ones. By contrast, Luvie did the opposite, choosing the container associated with the fruit that was the same as the one eaten by the trainer. For instance, upon seeing the trainer eat the apple, she went to the container which held the apple, and upon seeing him eat the banana, to the container which held the banana. Bert and Jessie responded in an intermediate fashion, choosing the container associated with food different from that which the trainer was seen to be eating after first choosing the opposite container for two and four trials, respectively.

Discussion Causal reasoning is difficult because a missing item must be reconstructed. While in simple learning, a monkey receives an electric shock when it presses a lever (and in observational learning observes that another receives an electric shock when it presses the lever), in causal reasoning the monkey does not observe the model press the lever but sees only its emotional response to the shock. Because in both simple and observational learning temporal contiguity between the lever press and electric shock are either experienced or observed, the relation between them is readily learned. Causal reasoning, however, does not provide such temporally contiguous events. While the monkey sees the relation between the lever press and the model's painful state, the chimpanzee does not actually see this relation, but experiences only the informant's emotional state, and must reconstruct from it the event that caused the state. In our tests, even though the chimpanzees experienced the same emotional states as a consequence of the same events, they were incapable of reconstructing those events from the emotional state of another chimpanzee. This helps clarify the striking difference in difficulty between ^e difference between these data and preliminary data reported earlier (Premack, 1983) comes from an unaccountable difference between our village and city children. Village children typically lagged city children by 6-12 months.

242

D. Premack, A. Premack

learning and reasoning, and suggests why the former is found in all species, the latter in exceedingly few. One might say that in the second experiment there is evidence for causal reasoning on the part of the children and perhaps one chimpanzee. This experiment can be seen as one of causal reasoning because here too there is a missing element to reconstruct. The subjects saw the trainer eating one or another food, but were never shown where he obtained it. Nevertheless, the children and perhaps one chimpanzee evidently did reconstruct the location of the food. What is most interesting about this outcome is that subjects "asked" the question of where the trainer got the food, and "answered" it quite specifically by going consistently to the container holding the food different from that eaten by the trainer. They assumed the food was the same as that which had been placed in the container, and in making this assumption believed the one container to be empty. It is not the specific content of the assumption alone that is of interest, but the additional fact that they made any assumption at all. Most species, we suggest, will make no assumptions. They will observe the trainer eating and never ask where he obtained the food. Such a question is asked only if one sees an event as part of a causal sequence in which there is a missing component. Could we induce our subjects to change their assumption? Suppose there is insufficient time for the trainer to recover the food placed in the containers. Would this affect choice? Keeping all other conditions constant, we tested this possibility by wrapping the fruit and pastries in elaborate packages before placing them in the containers. Now the trainer could not possibly have obtained the food from the containers - there was not sufficient time for him to unwrap these items. Children of 4 years and older were profoundly affected by this change. They no longer chose the container holding the pastry different from the one eaten by the trainer but chose at chance level between the two containers. By contrast, younger children and the chimpanzees were totally unaffected by the change in the availability of the item. They responded as before.

Labelling a causal action A causal action can be analysed into three components: the actor who carries out the action, the object on which he acts, and the instrument with which he performs the action. For instance, John paints a wall with a brush, cuts an apple with a knife, and washes his dirty socks in soapy water. Can young children and chimpanzees analyse such causal sequences? We devised a non-verbal procedure to answer this question by showing simple actions on a television monitor and giving our subjects markers that adhered to

Levels of causal understanding in chimpanzees and children

243

the screen and allowed them to identify each component of an action. A red square placed on John, for example, identified the actor; a green triangle on the apple, the recipient of the action; a blue circle on the knife, the instrument of the action. Sarah was given this problem. She was trained on three different actions: Bill cutting an orange, John marking a paper with a pencil, and Henry washing an apple with water. The trainer demonstrated the proper placement of each marker, handed the markers to Sarah, and corrected Sarah's errors. After reaching criterion on the three training tapes, she was given a transfer test in which all her responses were approved - our standard procedure in transfer tests. The tests were uniquely demanding, for the scenes were not merely new, but also decidedly more complex than those used in training. Where the training scenes had presented one person acting on one object with one instrument, the transfer scenes presented: two objects, only one of which was acted upon; two instruments, only one of which was used; and two persons, only one of whom carried out the action, the other being engaged in some scenes as an observer of the action of the first, and in other scenes as the recipient of action; for example, Bill brushed Bob's hair. Sarah passed the transfer tests, but at a relatively low level of accuracy. She was 85% correct in the use of the actor marker, 67% correct with the object marker, and 62% correct with the instrument marker. We attempted to improve her accuracy by training her on the transfer series, correcting her errors where previously we had approved all her responses. The attempt failed because she now placed the markers on the blank part of the screen (calling our attention to a fact we had previously overlooked - most of the screen is blank!). This tactic was entirely new, and brought the experiment to a halt. In retrospect, we recognize the experiment was needlessly difficult, and could have been simplified by dropping one of the categories, either of object or of instrument. Kim Dolgin (1981) as part of her doctoral research applied the same procedures to young children from 3.8 to 4.4 years, with an average age of 4 years, using the same non-verbal approach used with Sarah. The children failed the transfer tests. They did not properly identify the actor marker but drew a simpler distinction, animate/inanimate (or perhaps person/non-person). They reserved the actor marker for people but without regard for whether the person was an actor, observer, or recipient of the action. They made a similar simplification in the case of object and instrument markers, reserving them for non-persons, but without regard for whether the object or instrument was in actual use or simply present. With the children Dolgin took a further step not possible with Sarah. She told the children the meaning of each marker. For example, she presented the scene in which Bill cut the apple, and then showing the child the actor marker told her

244

D. Premack, A. Premack

"He's the one doing it", the object marker "This is what he's doing it to", and the instrument marker "This is what he's doing it with." The results were dramatic. The children passed the transfer tests at an average level of 93%, far higher than Sarah.

Causal sequences as transformations An actual causal sequence is a dynamic one in which an agent acts on an object, typically with the use of an instrument, changing its state and/or location. But one can represent the causal sequence in a stylized way - an apple and a cut apple representing the transformation of the apple, a knife the object responsible for the transformation. To determine whether non-speaking creatures can recognize such transformations, we designed a test in which the subject was given an incomplete representation of this causal sequence and was required to choose an alternative that properly completed it (Premack, 1976). The main actions we tested were cutting, marking, wetting and, in special cases, actions that reversed the effect: joining, erasing and drying. The subjects had extensive experience with the actions on which they were tested, carrying them out in a play context. The problem was given to three chimpanzees (and numerous children) in two basic formats, one requiring that they choose the missing operator, another the missing terminal state. In the former, they were given as a representative test sequence "apple ? cut apple" along with the alternatives: knife, pencil, container of water. In the latter, they were given as a representative test sequence "apple knife ?" along with the alternatives: cut apple, cut banana, wet apple. The chimpanzees were given not only novel object-operator combinations but also anomalous ones-for example, cut sponges, wet paper, fruit that had been written on - and performed as well on the anomalous material as on the other (see Premack, 1976, pp. 4-7, 249-261 for details). The tests were passed only by language-trained chimpanzees which were not given any training on the test but passed them on their first exposure to them. This test was only one of four that language-trained chimpanzees could do; the other three were analogies, same/different judgments on the relations between relations, and the matching of physically unlike proportions (e.g., i potato to \ glass of water) (Premack, 1983). Language training conferred an immediate ability tojsolve this complete set of four tasks. Only with protracted training could non-language-trained chimpanzees be taught to do these tests, and then only one test at a time, with no apparent saving from one to the other (Premack, 1988).

Levels of causal understanding in chimpanzees and children

245

Mapping the directionality of action The direction of the action was depicted in the test sequences by the left /right order of the objects. Thus, the object in its initial state was always presented on the left, the object in its transformed state on the right. But the standard tests did not require the subjects to read the sequences from left to right; they might have chosen an operator simply on the grounds that it belonged with a particular action-a knife, for example, as the operator for cutting - making this choice without regard to order. Whether the intact apple was to the left or right of the cut one, the animal could well have made the same choice. To obtain evidence that Sarah could discriminate the left-right order of the sequence and make use of it, she was first acquainted with pairs of actions that had reverse effects. For example, the trainer showed her how to mend broken pieces of an object with Scotch tape, how to erase pencil or crayon marks with a gum eraser, how to dry a wet object with a cloth-and then gave her the opportunity to carry out the actions. She adopted these new actions with enthusiasm. She was then given three test sessions with the original "cut, mark, wet", and four sessions with the new cases "tape, erase, dry". The tests took the standard form, for example, "apple ? cut apple", accompanied by the standard three operators, for example, knife, container of water, pencil. A total of 26 objects were used and 12 operators, two of each kind. Each of the three cases (tape, erase, dry) was presented four times per session in random order, with the position of the correct operator randomized across left, centre and right positions. Results: Total = 40/60 Original cases = 12/18 New cases = 28/42 Zdiff between old and new not significant. These preliminary results simply established that Sarah understood the new actions and could deal with them correctly. She was then required to use the left-right order of the sequence, and presented pairs of trials in which the same material appeared in reverse order. For example, "paper ? marked paper"; "marked paper ? paper". She was given pencil, container of water, and eraser as possible operators. Now, to choose correctly, Sarah had to take order into account, for while pencil was correct (and eraser incorrect) for one order, eraser was correct (and pencil incorrect) for the other. It takes a pencil to mark a blank sheet of paper, an eraser to remove the mark. Sarah was given 16 sessions, 12 trials per session, old cases being presented 24

246

D. Premack, A. Premack

times each, new cases 36 times each in random order across sessions. Although the objects and operators used were the same as those in the previous tests, they were combined in new ways. All other details were the same as those of the preceding tests. Results: Total = 110/180 Original cases = 47/72 New cases = 63/108 If we exclude trials in which Sarah chose the incorrect irrelevant alternative rather than the relevant one, corresponding figures are: Total = 110/148 Original cases = 47/62 New cases = 63/86 Zdiff between old and new not significant. The reversal pairs compared as follows: Cut/tape: 28/60 Mark /erase: 37/60 Wet/dry: 45/60 Zdiff between c/t and w/d = 3.18. Finally, Sarah was given an extensive transfer test involving new objects and operators. In five sessions of 12 trials per session, 30 new objects were used as well as 60 new operators, 10 of each kind. Each object appeared twice, once with one or another of the six new operators, and again with the reverse operator. Each operator appeared three times: as correct alternative, incorrect reverse alternative, and incorrect irrelevant alternative. Each case appeared twice per session in counterbalanced order. All other details were the same as those already reported. Results: Total = 44/60 Original cases = 20/30 New cases = 24/30 Excluding trials in which Sarah chose incorrect alternative: Total = 44/52 Original cases = 20/25

Levels of causal understanding in chimpanzees and children

247

New cases = 24/27 Reversal pairs: Cut/tape = 15/20 Mark/erase = 12/20 (p < .05 with three alternatives) Wet/dry = 17/20 No significant Zdiff. These data establish that Sarah could use left-right order to map the directionality of action as accurately on unfamiliar as on familiar cases.

Multiple transformations The basic consequence of a causal action is a transformation - a change from an initial state to a final one. Could Sarah understand causal action from this perspective, looking at the initial state of an object, comparing it with its terminal state, then selecting the operator(s) that explains or accounts for the difference? We can add to the interest of this question by removing the restrictions that were applied to the examples Sarah had been given. First, transformations involved more than a single action-for example, paper could be both cut and marked. Second, the initial state could be an already-transformed object rather than one in an intact or canonical state. Now we not only lifted restrictions, but gave Sarah a special trash bin in which to discard incorrect or irrelevant operators. So, besides removing the interrogative particle and replacing it with the correct or relevant operators, she was required to select the incorrect or irrelevant operators, and place it/them in the trash. The test consisted of six 12-trial sessions, each consisting of both single-action and double-action trials in equal number counterbalanced over the session. The six actions and their combinations were presented in equal number in each session counterbalanced over the session. The rest of the procedural details have already been reported. Results: Single transformations = 25/36 Double transformations = 24/36 (p < .001, both cases) Total = 49/72 These results add to the evidence of Sarah's ability to use the test sequences as representations of action. Her analyses answered these implicit questions: (1)

248

D. Premack, A. Premack

What operator changed the object from its initial to its terminal state? (2) In applying this operator to this object, what terminal state did one produce? (3) Which operators caused the difference between the initial and terminal state, and which did not? In answering these questions, Sarah had to attend to the order of the test sequences, "reading" them from left to right. We speculate that the representational ability required to pass these tests is that of a mind/brain which is capable of copying its own circuits. In carrying out an actual causal sequence, such as cutting an apple with a knife, an individual may form a neural circuit enabling him to carry out the act efficiently. But suppose he is not required to actually cut an apple, but is instead shown a representation of cutting - an incomplete depiction of the cutting sequence such as was given the chimpanzees - could he use the neural circuit to respond appropriately, that is, to complete the representation by choosing the missing element? Probably not, for the responses associated with the original circuit are those of actual cutting; they would not apply to repairing an incomplete representation of cutting. Moreover, the representation of a sequence can be distorted in a number of ways, not only by removing elements as in the chimpanzee test, but also by duplicating elements, misordering them, adding improper elements, or combinations of the above. To restore distorted sequences to their canonical form requires an ability to respond flexibly, for example, to remove elements, add others, restore order and the like. Flexible novel responding of this kind is not likely to be associated with the original circuitry (that concerned with actual cutting), but more likely with a copy of the circuit. Copies of circuits are not tied, as are the original circuits, to a fixed set of responses, and they may therefore allow for greater novelty and flexibility. For this reason, we suggest, flexible responding may depend on the ability of a mind/brain to be able to make copies of its own circuits.

An attempt to combine three questions Sarah was given a test that consisted of three questions: (1) What is the cause of this problem? (2) How can it be solved? (3) What is neither cause nor solution but merely an associate of the problem? These questions were not asked explicitly, with words, but implicitly with visual exemplars. Similarly, her answers were not given in words but in visual exemplars. The problems about which she was queried were depicted by a brief videotape (the terminal image of which was put on hold). For instance, she was shown a trainer vigorously stamping out a small conflagration of burning paper. The questions asked her in this case were: What caused the fire? How could it be put out? What is neither cause nor solution but an associate of the fire? The correct answers were photographs of: matches (cause), a bucket of water

Levels of causal understanding in chimpanzees and children

249

(solution), and a pencil (associate), the latter because she often used a pencil in scribbling on paper of exactly the kind that was shown in the videotape. The three questions were identified by different markers (like those used to identify the three components of an action), though the meaning of these markers was not determined by their location on the television image, but by the correct answers with which each marker was associated. The markers were introduced by presenting each of them with a videotape, offering three photographs and teaching her which of them was correct. In the example concerning a fire, which served as one of three training cases, she was presented the marker for "cause" (red square), with three photographsmatches, clay, knife-and taught to choose matches. When presented the marker for "solution" (green triangle), she was shown photographs of water, Scotch tape, eraser, and taught to choose the bucket of water. When presented the marker for associate (blue circle), she was shown photographs of pencil, apple, blanket, and taught to choose pencil. This procedure, repeated with two other training videotapes, was intended to teach her to view the problems depicted on the videotape according to the perspective indicated by the correct answer associated with a marker. Once she reached criterion on the three training cases, she was given a transfer test involving 20 new problems. When she failed, it was decided to train her on this material and to bring in a new trainer - one who no longer played an active role in her care or testing but who had been an early caretaker, was a favourite, and could be counted on to bring out her best effort. Sarah definitely "tried" harder with some trainers than with others. Ordinarily she looked only once at an alternative before choosing, but with a difficult question and a favourite trainer, several times. Her looking behaviour was readily observable; after the trainer gave her the test material in a manila envelope, he left the cage area. Sarah could be observed on a television monitor to empty the envelope on the cage floor, spread out the alternatives, inspect them, choose, and then ring her bell (as a period marker signalling an end and summoning the trainer). With this favourite trainer she not only looked at the alternatives with more than usual care but did several double-takes; that is, looked, looked away, and then quickly looked back. This did not help her cause; she made three consecutive errors. When the trainer entered to show her the fourth videotape, she lost sphincter control and ran screaming about the cage. Although a demanding test, it is not necessarily beyond the capacity of the chimpanzee. It must be taught more carefully than we did, not as a combination of three markers, but one marker at a time, then two and, only when there is success on two, all three presented together. We subsequently used this approach with 4^-6^-year-old children on a problem only slightly less demanding than the one given Sarah, and they succeeded nicely (Premack, 1988).

250

D. Premack, A. Premack

General discussion There are two traditions in which causality has been studied in psychology: the natural causality of Michotte (1963), and the arbitrary causality of Hume (1952). Natural causality (Premack, 1991) concerns the relation between special pairs of events-one object launching another by colliding with it is the classic example; whereas arbitrary causality concerns the relation between any pair of temporally contiguous events - a lever press followed by the presentation of a Mars bar is one example. These two traditions have fostered conflicting interpretations of causality. The perception of natural causal relations requires only a single pairing of appropriate events, is demonstrable in 6-month-old infants (e.g., Leslie & Keeble, 1987) and is considered innate; whereas the perception of arbitrary causal relations requires repeated pairings in time of two events, and is learned. But are these differences real or do they simply reflect differences in subject matter? Although Michotte's case is typically the only cited example of natural causality, it is essential to recognize that there is another, and more basic example of natural causality. This is the psychological case where we perceive causality under two conditions: (1) when an object moves without being acted upon by another object, that is, is "self-propelled" (Premack, 1990); and (2) when one object affects another "at a distance", that is, affects another despite a lack of spatial/temporal contiguity. Humans unquestionably perceive causality under both these conditions. Yet this fact has received little comment - virtually none compared to the extensive comment on the Michotte case. Why? We suggest, because the perception of causality in the psychological domain evolved far earlier than it did in the Michotte case and belongs to a "part of the mind" that is less accessible to language. The perception of causality of the Michotte variety probably evolved late, and even then only in a few tool-using species (Kummer, in press), for only humans, apes, and the exceptional monkey (e.g., Cebus) handle objects in a manner capable of producing collisions. Compare this to the perception of causality in the psychological domain, which is not restricted to a few tool-using species but is found in virtually all species. Intentional action which involves either a single individual or one individual acting upon another is part of the experience of all but a few invertebrates. Bar pressing that produces food, threats that produce withdrawal, collisions that launch objects, and so on, all fit neatly into either a physical or psychological category. These cases are important because they give the impression that the concept of cause has a content: psychological, physical, or both. However, infants may perceive a causal relationship when presented with totally arbitrary cases; for

Levels of causal understanding in chimpanzees and children

251

example, a sharp sound followed by a temporally contiguous colour change in an object. If so, this example is important because it demonstrates that the concept of cause may be without content. Let us habituate one group of infants to a contiguous case, another to a delay case, and then apply the Leslie-Keeble paradigm by reversing the order of the two events. We present colour change followed by sound to both groups. The contiguous group, we predict, will show greater dishabituation (recovery in looking time) than the delay group. In other words, our outcome will be the same as that obtained by LeslieKeeble in the Michotte example - a greater recovery for the group in which the events are presented contiguously. But how does one explain these results? Just as Leslie-Keeble do. While contiguous events certainly do lead to the perception of causality, is this perception confined to the Michotte case? Causality is not bound by content, we suggest, for the sequence "sound-colour change" which involves neither an intentional act nor the transfer of energy from one object to another, is an example of neither psychological nor physical causality. Perhaps the concept of causality at its most fundamental level is no more than a device that records the occurrence of contiguous events - associative learning - and is found in all species. All species may share a device that records the occurrence of contiguous events, and evolution contributed two major additions to this primitive device: first, the capacity to act intentionally which enabled certain species not only to register but also to produce contiguous events; second, the capacity, largely unique to the human, to explain or interpret events that have been both registered and produced. While the basis of the primitive level has not been resolved by neuroscience (for this level operates on "events", and how the mind/brain binds items so as to construct events remains a challenge for neuroscience; e.g., Singer, 1990), fortunately, the second level of causality is well represented by work on animal learning. Especially Dickinson and his colleagues (e.g., Dickinson & Shanks, in press) have considered the special subset of concurrent events - act-outcome pairs - brought about by intentional action. Are such pairs marked in some fashion, and thus represented differently in memory from other concurrent pairs? The third level of causality is to be found in recent work on domain-specific theories of naive physics (Spelke, Phillips, & Woodward, in press; Baillargeon, Kotovsky, & Needham, in press), psychology (Leslie, in press; Gelman, Durgin, & Kaufman, in press; Premack & Premack, in press), and arguably biology (Carey, in press; Keil, in press). These theories separate the concurrent events (registered by the primitive device) into special categories, and propose an explanatory mechanism for each of them. Explanation, embedded in naive theories about the world, is largely a human specialization.

252

D. Premack, A. Premack

References Baillargeon, R., Kotovsky, L., & Needham, A. (in press). The acquisition of physical knowledge in infancy. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Carey, S. (in press). On the origin of causal understanding. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Dickinson, A. & Shanks, D. (in press). Instrumental action and causal representation. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Dolgin, K.G. (1981). A developmental study of cognitive predisposition'. A study of the relative salience of form and function in adult and four-year-old subjects. Dissertation, University of Pennsylvania. Gelman, R., Durgin, F., & Kaufman, L. (in press). Distinguishing between animates and inanimates: Not by motion alone. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Gillan, D.J. (1981). Reasoning in the chimpanzee. II. Transitive inference. Journal of Experimental Psychology: Animal Behavior Processes, 7, 150-164. Gillan, D.J., Premack, D., & Woodruff, G. (1981). Reasoning in the chimpanzee. I. Analogical reasoning. Journal of Experimental Psychology: Animal Behavior Processes, 7, 1-17. Hume, D. (1952). An enquiry concerning human understanding. In Great books of the western world (Vol. 35). Chicago: Benton. Keil, F. (in press). The growth of causal understandings of natural kinds. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Kummer, H. (in press). On causal knowledge in animals. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Leslie, A., & Keeble, S. (1987). Do six-month-old infants perceive causality? Cognition, 25, 265-288. Michotte, A. (1963). The perception of causality. London: Methuen. Premack, D. (1976). Intelligence in ape and man. Hillsdale, NJ: Erlbaum. Premack, D. (1983). The codes of man and beasts. Behavioral and Brain Sciences, 6, 125-167. Premack, D. (1988). Minds with and without language. In L. Weiskrantz (Ed.), Thought without language. Oxford: Clarendon Press. Premack, D. (1990). The infant's theory of self-propelled objects. Cognition, 36, 1-16. Premack, D. (1991). Cause/induced motion: intention/spontaneous motion. Talk at Fyssen Foundation Conference on the origins of the human brain. Clarendon Press: Oxford. Premack, D., & Premack, A.J. (in press). Semantics of action. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press. Singer, W. (1990). Search for coherence: A basic principle of cortical self-organization. Concepts in Neuroscience, 1, 1-26. Spelke, E.S., Phillips, A., & Woodward, A.L. (in press). Infant's knowledge of object motion and human action. In A.J. Premack, D. Premack, & D. Sperber (Eds.), Causal cognition: A multidisciplinary debate. Oxford: Clarendon Press.

13 Uncertainty and the difficulty of thinking through disjunctions Eldar Shafir* Department of Psychologyy Princeton University, Princeton, NJ 08544, USA

Abstract This paper considers the relationship between decision under uncertainty and thinking through disjunctions. Decision situations that lead to violations of Savage's sure-thing principle are examined, and a variety of simple reasoning problems that often generate confusion and error are reviewed. The common difficulty is attributed to people's reluctance to think through disjunctions. Instead of hypothetically traveling through the branches of a decision tree, it is suggested, people suspend judgement and remain at the node. This interpretation is applied to instances of decision making, information search, deductive and inductive reasoning, probabilistic judgement, games, puzzles and paradoxes. Some implications of the reluctance to think through disjunctions, as well as potential corrective procedures, are discussed.

Introduction Everyday thinking and decision making often occur in situations of uncertainty. A critical feature of thinking and deciding under uncertainty is the need to consider possible states of the world and their potential consequences for our beliefs and actions. Uncertain situations may be thought of as disjunctions of possible states: either one state will obtain, or another. In order to choose between alternative actions or solutions in situations of uncertainty, a person

* E-mail eidar@clarity. princeton. edu This research was supported by US Public Health Service Grant No. 1-R29-MH46885 from the National Institute of Mental Health, and by a grant from the Russell Sage Foundation. The paper has benefited from long discussions with Amos Tversky, and from the comments of Philip Johnson-Laird, Daniel Osherson, and Edward Smith.

254

E. Shafir

needs to consider the anticipated outcomes of each action or each solution pattern under each state. Thus, when planning a weekend's outing, a person may want to consider which of a number of activities she would prefer if the weekend is sunny and which she would prefer if it rains. Similarly, when contemplating the next move in a chess game, a player needs to consider what the best move would be if the opponent were to employ one strategy, and what may be the best move if the opponent were to follow an alternative plan. Special situations sometimes arise in which a particular action, or solution, yields a more desirable outcome no matter how the uncertainty is resolved. Thus, a person may prefer to go bowling rather than hiking regardless of whether it is sunny or it rains, and an exchange of queens may be the preferred move whatever the strategy chosen by the opponent. An analogous situation was described by Savage (1954) in the following passage: A businessman contemplates buying a certain piece of property. He considers the outcome of the next presidential election relevant to the attractiveness of the purchase. So, to clarify the matter for himself, he asks whether he would buy if he knew that the Republican candidate were going to win, and decides that he would do so. Similarly, he considers whether he would buy if he knew that the Democratic candidate were going to win, and again finds that he would do so. Seeing that he would buy in either event, he decides that he should buy, even though he does not know which event obtains . . .

Savage calls the principle that governs this decision the sure-thing principle (STP). According to STP, if a person would prefer a to b knowing that X obtained, and if he would also prefer a to b knowing that X did not obtain, then he definitely prefers a to b (Savage, 1954, p. 22). STP has a great deal of both normative and descriptive appeal, and is one of the simplest and least controversial principles of rational behavior. It is an important implication of "consequentialist" accounts of decision making, in that it captures a fundamental intuition about what it means for a decision to be determined by the anticipated consequences.1 It is a cornerstone of expected utility theory, and it holds in other models of choice which impose less stringent criteria of rationality (although see McClennen, 1983, for discussion). Despite its apparent simplicity, however, people's decisions do not always abide by STP. The present paper reviews recent experimental studies of decision under uncertainty that exhibit violations of STP in simple disjunctive situations. It is argued that a necessary condition for such violations is people's failure to see through the underlying disjunctions. In particular, it is suggested that in situations of uncertainty people tend to refrain from fully contemplating the consequences of potential outcomes and, instead, suspend judgement and remain, undecided, at ^he notion of consequentialism appears in the decision theoretic literature in a number of different senses. See, for example, Hammond (1988), Levi (1991), and Bacharach and Hurley (1991) for technical discussion. See also Shafir and Tversky (1992) for a discussion of nonconsequentialism.

Uncertainty and the difficulty of thinking through disjunctions

255

the uncertain node. Studies in other areas, ranging from deduction and probability judgement to games and inductive inference, are then considered, and it is argued that a reluctance to think through disjunctions can be witnessed across these diverse domains. Part of the difficulty in thinking under uncertainty, it is suggested, derives form the fact that uncertainty requires thinking through disjunctive situations. Some implications and corrective procedures are considered in a concluding section.

Decisions Risky choice Imagine that you have just gambled on a toss of a coin in which you had an equal chance to win $200 or lose $100. Suppose that the coin has been tossed, but that you do not know whether you have won or lost. Would you like to gamble again, on a similar toss? Alternatively, how would you feel about taking the second gamble given that you have just lost $100 on the first (henceforth, the Lost version)? And finally, would you play again given that you have won $200 on the first toss (the Won version)? Tversky and Shafir (1992) presented subjects with the Won, Lost, and uncertain versions of this problem, each roughly a week apart. The problems were embedded among several others so the relation among the three versions would not be transparent, and subjects were instructed to treat each decision separately. The data were as follows: the majority of subjects accepted the second gamble after having won the first gamble, the majority accepted the second gamble after having lost the first gamble, but most subjects rejected the second gamble when the outcome of the first was not known. Among those subjects who accepted the second gamble both after a gain and after a loss on the first, 65% rejected the second gamble in the disjunctive condition, when the outcome of the first gamble was uncertain. In fact, this particular pattern - accept when you win, accept when you lose, but reject when you do not know - was the single most frequent pattern exhibited by our subjects (see Tversky & Shafir, 1992, for further detail and related data). A decision maker who would choose to accept the second gamble both after having won and after having lost the first, s h o u l d - i n conformity with S T P choose to accept the second gamble even when the outcome of the first is uncertain. However, when it is not known whether they have won or lost, our subjects refrain from contemplating (and acting in accordance with) the consequences of winning or of losing. Instead, they act as if in need for the uncertainty about the first toss to be resolved. Elsewhere, we have suggested that people have

256

E. Shafir

different reasons for accepting the second gamble following a gain and following a loss, and that a disjunction of different reasons (" 'I can no longer lose. . .' in case I won the first gamble or 1 need to recover my losses. . .' in case I lost") is often less compelling than either of these definite reasons alone (for further discussion of the role of reasons in choice, see Shafir, Simonson, & Tversky, 1993). Tversky and Shafir (1992) call the above pattern of decisions a disjunction effect. A disjunction effect occurs when a person prefers x over y when she knows that event A obtains, and she also prefers x over y when she knows that event A does not obtain, but she prefers y over x when it is unknown whether or not A obtains. The disjunction effect amounts to a violation of STP, and hence of consequentialism. While a reliance on reasons seems to play a significant role in the psychology that yields disjunction effects, there is nonetheless another important element that contributes to these paradoxical results: people do not see through the otherwise compelling logic that characterizes these situations. When confronting such disjunctive scenarios, which can be thought of as decision trees, people seem to remain at the uncertain nodes, rather then contemplate t h e sometimes incontrovertible - consequences of the possible branches. The above pattern of nonconsequential reasoning may be illustrated with the aid of the value function from Kahneman and Tversky's (1979) prospect theory. The function, shown in Fig. 1, represents people's subjective value of losses and of gains, and captures common features of preference observed in numerous empirical studies. Its S-shape combines a concave segment to the right of the origin reflecting risk aversion in choices between gains, and a convex segment to the left of the origin reflecting risk seeking in choices between losses. Furthermore, the slope of the function is steeper on the left of the origin than on the right, reflecting the common observation that "losses loom larger than gains" for most people. (For more on prospect theory, see Kahneman & Tversky, 1979, 1982, as well as Tversky & Kahneman, 1992, for recent extensions.) The function in Fig. 1 represents a typical decision maker who is indifferent between a 50% chance of winning $100 and a sure gain of roughly $35, and, similarly, is indifferent between a 50% chance of losing $100 and a sure loss of roughly $40. Such a pattern of preferences can be captured by a power function with an exponent of .65 for gains and .75 for losses. While prospect theory also incorporates a decision weight function, 7r, which maps stated probabilities into their subjective value for the decision maker, we will assume, for simplicity, that decision weights coincide with stated probabilities. While there is ample evidence to the contrary, this does not change the present analysis. Consider, then, a person P whose values for gains and losses are captured by the function of Fig. 1. Suppose that P is presented with the gamble problem above and is told that he has won the first toss. He now needs to decide whether to accept or reject the second. P needs to decide, in other words, whether to

Uncertainty and the difficulty of thinking through disjunctions

Fig. 1.

257

The value function v(x) = x65 for x^O and v(x) = -(-x)m75 for x^O.

maintain a sure gain of $200 or, instead, opt for an equal chance at either a $100 or a $400 gain. Given P's value function, his choice is between two options whose expected values are as follows: Accept the second gamble: Reject the second gamble:

.50 x 400( 65) + .50 x 100( 65) 1.0 x 200( 65)

Because the value of the first option is greater than that of the second, P is predicted to accept the second gamble. Similarly, when P is told that he has lost the first gamble and needs to decide whether to accept or reject the second, P faces the following options: Accept the second gamble: Reject the second gamble:

.50 x -[200 ( 75)] + .50 x 100( 65) 1.0 x -[100 ( 75)]

258

E. Shafir

Again, since the first quantity is larger than the second, P is predicted to accept the second gamble. Thus, once the outcome of the first gamble is known, the value function of Fig. 1 predicts that person P will accept the second gamble whether he has won or lost the first. But what is P expected to do when the outcome of the first gamble is not known? Because he does not know the outcome of the first gamble, P may momentarily assume that he is still where he began-that, for moment, no changes have transpired. Not knowing whether he has won or lost, P remains for now at the status quo, at the origin of his value function. When presented with the decision to accept or reject the second gamble, P evaluates it from his original position, without incorporating the outcome of the first gamble, which remains unknown. Thus, P needs to choose between accepting or rejecting a gamble that offers an equal chance to win $200 or lose $100: Accept the second gamble: .50 x -[100 ( 75)] + .50 x 200( 65) Reject the second gamble: 0 Because the expected value of accepting is just below 0, P decides to reject the second gamble in this case. Thus, aided by prospect theory's value function, we see how a decision maker's "suspension of judgement" - his tendency to assume himself at the origin, or status quo, when it is not known whether he has won or lost - leads him to reject an option that he would accept no matter what his actual position may be. Situated at a chance node whose outcome is not known, P's reluctance to consider each of the hypothetical branches leads him to behave in a fashion that conflicts with his preferred behavior given either branch. People in these situations seem to confound their epistemic uncertainty - what they may or may not know-with uncertainty about the actual consequences - what may or may not have occurred. A greater focus on the consequences would have helped our subjects realize the implications for their preference of either of the outcomes. Instead, not knowing which was the actual outcome, our subjects chose to evaluate the situation as if neither outcome had obtained. It is this reluctance to think through disjunctions that characterizes many of the phenomena considered below.

Search for noninstrumental information: the Hawaiian vacation Imagine that you have just taken a tough qualifying exam. It is the end of the semester, you feel tired and run-down, and you are not sure that you passed the exam. In case you failed you have to take it again in a couple of months-after

Uncertainty and the difficulty of thinking through disjunctions

259

the Christmas holidays. You now have an opportunity to buy a very attractive 5-day Christmas vacation package to Hawaii at an exceptionally low price. The special offer expires tomorrow, while the exam grade will not be available until the following day. Do you buy the vacation package? This question was presented by Tversky and Shafir (1992) to Stanford University undergraduate students. Notice that the outcome of the exam will be known long before the vacation begins. Thus, the uncertainty characterizes the present, disjunctive situation, not the eventual vacation. Additional, related versions were presented in which subjects were to assume that they had passed the exam, or that they had failed, before they had to decide about the vacation. We discovered that many subjects who would have bought the vacation to Hawaii if they were to pass the exam and if they were to fail, chose not to buy the vacation when the exam's outcome was not known. The data show that more than half of the students chose the vacation package when they knew that they passed the exam and an even larger percentage chose the vacation when they knew that they failed. However, when they did not know whether they had passed or failed, less than one-third of the students chose the vacation and the majority (61%) were willing to pay $5 to postpone the decision until the following day, when the results of the exam would be known.2 Note the similarity of this pattern to the foregoing gamble scenario: situated at a node whose outcome is uncertain, our students envision themselves at the status quo, as if no exam had been taken. This "suspension of judgement" - the reluctance to consider the possible branches (having either passed or failed the exam) - leads our subjects to behave in a manner that conflicts with their preferred option given either branch. The pattern observed in the context of this decision is partly attributed by Tversky and Shafir (1992) to the reasons that subjects summon for buying the vacation (see also Shafir, Simonson, & Tversky, 1993, for further discussion). Once the outcome of the exam is known, the student has good - albeit different reasons for going to Hawaii: having passed the exam, the vacation can be seen as a reward following a successful semester; having failed the exam, the vacation becomes a consolation and time to recuperate before a re-examination. Not knowing the outcome of the exam, however, the student lacks a definite reason for going to Hawaii. The indeterminacy of reasons discourages many students from buying the vacation, even when both outcomes - passing or failing the exam - ultimately favor this course of action. Evidently, a disjunction of different

2 Another group of subjects were presented with both Fail and Pass versions, and asked whether they would buy the vacation package in each case. Two-thirds of the subjects made the same choice in the two conditions, indicating that the data for the disjunctive version cannot be explained by the hypothesis that those who buy the vacation in case they pass the exam do not buy it in case they fail, and vice versa. While only one-third of the subjects made different decisions depending on the outcome of the exam, more than 60% of the subjects chose to wait when the outcome was not known.

260

E. Shafir

reasons (reward in case of success; consolation in case of failure) can be less compelling than either definite reason alone. A significant proportion of subjects were willing to pay, in effect, for information that was ultimately not going to alter their decision - they would choose to go to Hawaii in either case. Such willingness to pay for noninstrumental information is at variance with the classical model, in which the worth of information is determined by its potential to influence choice. People's reluctance to think through disjunctive situations, on the other hand, entails that noninstrumental information will sometimes be sought. (See Bastardi & Shafir, 1994, for additional studies of the search for noninstrumental information and its effects on choice.) While vaguely aware of the possible outcomes, people seem reluctant to fully entertain the consequences as long as the actual outcome is uncertain. When seemingly relevant information may become available, they often prefer to have the uncertainty resolved, rather than consider the consequences of each branch of the tree under the veil of uncertainty. A greater tendency to consider the potential consequences may sometimes help unveil the noninstrumental nature of missing information. In fact, when subjects were first asked to contemplate what they would do in case they failed the exam and in case they passed, almost no subject who had expressed the same preference for both outcomes then chose to wait to find out which outcome obtained (Tversky & Shafir, 1992). The decision of many subjects in the disjunctive scenario above was not guided by a simple evaluation of the consequences (for, then, they would have realized that they prefer to go to Hawaii in either case). An adequate account of this behavior needs to contend with the fact that the very simple and compelling disjunctive logic of STP does not play a decisive role in subjects' reasoning. A behavioral pattern which systematically violates a simple normative rule requires both a positive as well as a negative account (see Kahneman and Tversky, 1982, for discussion). We need to understand not only the factors that produce a particular response, but also why the correct response is not made. Work on the conjunction fallacy (Shafir, Smith, & Osherson, 1990; Tversky and Kahneman, 1983), for example, has addressed both the fact that people's probability judgement relies on the representativeness heuristic - a positive account - as well as the fact that people do not perceive the extensional logic of the conjunction rule as decisive - a negative account. The present work focuses on the negative facet of nonconsequential reasoning and STP violations. It argues that like other principles of reasoning and decision making, STP is very compelling when stated in a general and abstract form, but is often non-transparent, particularly because it applies to disjunctive situations. The following section briefly reviews studies of nonconsequential decision making in the context of games, and ensuing sections extend the analysis to other domains.

Uncertainty and the difficulty of thinking through disjunctions

261

Games Prisoner's dilemma The theory of games explores the interaction between players acting according to specific rules. One kind of two-person game that has received much attention is the Prisoner's dilemma, or PD. (For an extensive treatment, see Rapoport & Chammah, 1965). A typical PD is presented in Fig. 2. The cell entries indicate the number of points each player receives contingent on the two players' choices. Thus, if both cooperate each receives 75 points but if, for example, the other cooperates and you compete, you receive 85 points while the other receives 25. What characterizes the PD is that no matter what the other does, each player fares better if he competes than if he cooperates; yet, if they both compete they do significantly less well than if they had both cooperated. Since each player is encountered at most once, there is no opportunity for conveying strategic messages, inducing reciprocity, or otherwise influencing the other player's choice of strategy. A player in a PD faces a disjunctive situation. The other chooses one of two strategies, either to compete or to cooperate. Not knowing the other's choice, the first player must decide on his own strategy. Whereas each player does better competing, their mutually preferred outcome results from mutual cooperation rather than competition. A player, therefore, experiences conflicting motivations. Regardless of what the other does, he is better off being selfish and competing; but assuming that the other acts very much like himself, they are better off both making the ethical decision to cooperate rather than the selfish choice to compete. How might this disjunctive situation influence people's choice of strategy?

OTHER cooperates

competes

You: 75

You: 25

Other: 75

Other 85

You: 85

You: 30

Other 25

Other 30

cooperate

YOU

compete

— Fig. 2.

1

A typical prisoner's dilemma. The cell entries indicate the number of points that you and the other player receive contingent on your choices.

262

E. Shafir

Shafir and Tversky (1992) have documented disjunction effects in one-shot PD games played for real payoffs. Subjects (N = 80) played a series of PD games (as in Fig. 2) on a computer, each against a different unknown opponent supposedly selected at random from among the participants. Subjects were told that they had been randomly assigned to a "bonus group", and that occasionally they would be given information about the other player's already-chosen strategy before they had to choose their own. This information appeared on the screen next to the game, and subjects were free to take it into account in making their decision. (For details and the full instructions given to subjects, see Shafir & Tversky, 1992.) The rate of cooperation in this setting was 3% when subjects knew that the opponent had defected, and 16% when they knew that the opponent had cooperated. Now what should subjects do when the opponent's decision is not known? Since 3% cooperate when the other competes and 16% cooperate when the other cooperates, one would expect an intermediate rate of cooperation when the other's strategy is not known. Instead, when subjects did not know whether their opponent had cooperated or defected (as is normally the case in this game), the rate of cooperation rose to 37%. In violation of STP, a quarter of the subjects defected when they knew their opponent's choice-be it cooperation or defection - but cooperated when their opponent's choice was not known. Note the recurring pattern: situated at a disjunctive node whose outcome is uncertain, these subjects envision themselves at the status quo, as if, for the moment, the uncertain strategy selected by the opponent has no clear consequences. These players seem to confound their epistemic uncertainty - what they may or may not know about the other's choice of strategy - with uncertainty about the actual consequences - the fact that the other is bound to be a cooperator or a defector, and that they, in turn, are bound to respond by defecting in either case. (For further analysis and a positive account of what may be driving subjects' tendency to cooperate under uncertainty, see Shafir & Tversky, 1992.)

Newcomb's problem and quasi-magical thinking Upon completing the PD game described in fhe previous section, subjects (N = 40) were presented, on a computer screen, with the following scenario based on the celebrated Newcomb's problem (for more on Newcomb's problem, see Nozick, 1969; see Shafir & Tversky, 1992, for further detail and discussion of the experiment). You now have one more chance to collect additional points. A program developed recently at MIT was applied during this entire session to analyze the pattern of your preferences. Based on that analysis, the program has predicted your preference in this final problem.

Uncertainty and the difficulty of thinking through disjunctions

I 20 points J Box A

|

?

263

I

Box B

Consider the two boxes above. Box A contains 20 points for sure. Box B may or may not contain 250 points. Your options are to: (1) Choose both boxes (and collect the points that are in both). (2) Choose Box B only (and collect only the points that are in Box B). If the program predicted, based on observation of your previous preferences, that you will take both boxes, then it left Box B empty. On the other hand, if it predicted that you will take only Box B, then it put 250 points in that box. (So far, the program has been remarkably successful: 92% of the participants who choose only Box B found 250 points in it, as opposed to 17% of those who chose both boxes.) To insure that the program does not alter its guess after you have indicated your preference, please indicate to the person in charge whether you prefer both boxes or Box B only. After you indicate your preference, press any key to discover the allocation of points.

According to one rationale that arises in the context of this decision, if the person chooses both boxes, then the program, which is remarkably good at predicting preferences, is likely to have predicted this and will not have put the 250 points in the opaque box. Thus, the person will get only 20 points. If, on the other hand, the person takes only the opaque box, the program is likely to have predicted this and will have put the 250 points in that box, and so the person will get 250 points. A subject may thus be tempted to reason that if he takes both boxes he is likely to get only 20 points, but that if he takes just the opaque box he is likely to get 250 points. There is a compelling motivation to choose just the opaque box, and thereby resemble those who typically find 250 points it it. There is, of course, another rationale: the program has already made its prediction and has already either put the 250 points in the opaque box or has not. If it has already put the 250 points in the opaque box, and the person takes both boxes he gets 250 + 20 points, whereas if he takes only the opaque box, he gets only 20 points. If the program has not put the 250 points in the opaque box and the person takes both boxes he gets 20 points, whereas if he takes only the opaque box he gets nothing. Therefore, whether the 250 points are there or not, the person gets 20 points more by taking both boxes rather than the opaque box only. The second rationale relies on consequentialist reasoning reminiscent of STP (namely, whatever the state of the boxes following the program's prediction, I will do better choosing both boxes rather than one only). The first rationale, on the other hand, while couched in terms of expected value, is partially based on the assumption that what the program will have predicted - although it has predicted this already - depends somehow on what the subject ultimately decides to do. The results we obtained were as follows: 35% of the subjects chose both boxes, while 65% preferred to take Box B only. This proportion of choices is similar to that observed in other surveys concerning the original Newcomb's problem (see,

264

E. Shafir

for example, Gardner, 1973, 1974; Hofstadter, 1983). What can be said about the majority who prefer to take just one box? Clearly, had they known for certain that there were 250 points in the opaque box (and could see 20 in the other), they would have taken both rather than just one. And certainly, if they knew that the 250 points were not in that box, they would have taken both rather than just the one that's empty. These subjects, in other words, would have taken both boxes had they known that Box B is either full or empty, but a majority preferred to take only Box B when its contents were not known. The conflicting intuitions that subjects experience in the disjunctive situation when the program's prediction is not known - are obviously resolved in favor of both boxes once the program's decision has been announced: at that point, no matter what the program has predicted, taking both boxes brings more points. Subjects, therefore, should choose both boxes also when the program's decision is uncertain. Instead, many subjects fail to be moved by the foreseeable consequences of the program's predictions, and succumb to the strong motivation to choose just the opaque box and thereby resemble those who typically find 250 points in it.3 As Gibbard and Harper (1978) suggest in an attempt to explain people's choice of a single box, "a person may. . . want to bring about an indication of a desired state of the world, even if it is known that the act that brings about the indication in no way brings about the desired state itself. This form of magical thinking was demonstrated by Quattrone and Tversky (1984), whose subjects selected actions that were diagnostic of favorable outcomes even though the actions could not cause those outcomes. Note that such instances of magical thinking typically occur in disjunctive situations, before the exact outcome is known. Once they are aware of the outcome, few people think they can reverse it by choosing an action that is diagnostic of an alternative event. Shafir and Tversky (1992) discuss various manifestations of "quasi-magical" thinking, related to phenomena of self-deception and illusory control. These include people's tendency to place larger bets before rather than after a coin has been tossed (Rothbart & Snyder, 1970; Strickland, Lewicki, & Katz, 1966), or to throw dice softly for low numbers and harder for high ones (Henslin, 1967). Similarly, Quattrone and Tversky (1984) note that Calvinists act as if their behavior will determine whether they will go to heaven or to hell, despite their belief in divine pre-determination, which entails that their fate has been determined at birth. The presence of uncertainty, it appears, is a major contributor to quasi-magical thinking; few people act as if they can undo an already certain 3 The fact that subjects do not see through this disjunctive scenario seems indisputable. It is less clear, however, what conditions would serve to make the situation more transparent, and to what extent. Imagine, for example, that subjects were given a sealed copy of the program's decision to take home with them, and asked to inspect it that evening, after having made their choice. It seems likely that an emphasis on the fact that the program's decision has long been made would reduce the tendency to choose a single box.

Uncertainty and the difficulty of thinking through disjunctions

265

event, but while facing a disjunction of events, people often behave as if they can exert some control over the outcome. Thus, many people who are eager to vote while the outcome is pending, may no longer wish to do so once the outcome of the elections has been determined. In this vein, it is possible that Calvinists would perhaps do fewer good deeds if they knew that they had already been assigned to heaven, or to hell, than while their fate remains a mystery. Along similar lines, Jahoda (1969) discusses the close relationship between uncertainty and superstitious behavior, which is typically exhibited in the context of uncertain outcomes rather than in an attempt to alter events whose outcome is already known. As illustrated by the studies above, people often are reluctant to consider the possible outcomes of disjunctive situations, and instead suspend judgement and envision themselves at the uncertain node. Interestingly, it appears that decision under uncertainty is only one of numerous domains in which subjects exhibit a reluctance to think through disjunctive situations. The difficulties inherent to thinking through uncertainty and, in particular, people's reluctance to think through disjunctions manifest themselves in other reasoning and problem-solving domains, some of which are considered below.

Probabilistic judgement Researchers into human intuitive judgement as well as teachers of statistics have commented on people's difficulties in judging the probabilities of disjunctive events (see, for example, Bar-Hillel, 1973; Carlson & Yates, 1989; Tversky & Kahneman, 1974). While some disjunctive predictions may in fact be quite complicated, others are simple, assuming that one sees though their disjunctive character. Consider, for example, the following "guessing game" which consisted of two black boxes presented to Princeton undergraduates (N = 40) on a computer screen, along with the following instructions. Under the black cover, each of the boxes above is equally likely to be either white, blue, or purple. You are now offered to play one of the following two games of chance: Game 1: You guess the color of the left-hand box. You win 50 points if you were right, and nothing if you were wrong. Game 2: You choose to uncover both boxes. You win 50 points if they are the same color, and nothing if they are different colors.

The chances of winning in Game 1 are 1/3; the chances of winning in Game 2 are also 1/3. To see that, one need only realize that the first box is bound to be either white, blue, or purple and that, in either case, the chances that the other will be the same color are 1/3. Notice that this reasoning incorporates the disjunctive

266

E. Shafir

logic of STP. One enumerates the possible outcomes of the first box, considers the chances of winning conditional on each outcome, and realizes that the chances are the same no matter what the first outcome was. Subjects, therefore, are expected to find the two games roughly equally attractive, provided that they see through the disjunctive nature of Game 2. This disjunctive rationale, however, seems not to have been entirely transparent to our subjects, 70% of whom indicated a preference for Game 1 (significantly different from chance, Z = 2.53, p