The Enigma of Reason: A New Theory of Human Understanding

by Dan Sperber

  The Infant, the Caterpillar, and the Hidden Piece of Cheese

  Metarepresentations, that is, representations of representations, became a major topic of study in psychology after David Premack and Guy Woodruff asked in a famous 1978 article, “Does the Chimpanzee Have a Theory of Mind?”3 The phrase “theory of mind” made for an attention-grabbing title, but it also became a source of theoretical misunderstandings. Premack and Woodruff’s question was not really whether chimpanzees have theoretical beliefs about minds (or, in our terms, use representations of psychological regularities as premises in inference). It was rather whether chimpanzees are capable of attributing specific beliefs or intentions to each other (or to humans).

  Some authors, such as Alison Gopnik and Henry Wellman or Josef Perner, believe that you need some theoretical understanding of mental states to attribute mental states to others.4 Others authors, such as Renée Baillargeon and Alan Leslie, don’t—and we agree with them.5 To avoid the confusion caused by the phrase “theory of mind,” we will use the metaphor mindreading to describe the cognitive ability involved.

  That humans are capable of mindreading is all too obvious. We attribute mental representations to one another all the time. We are often aware of what people around us think, and even of what they think we think. Such thoughts about the thoughts of others come to us quite naturally.

  There is no evidence, on the other hand, that most animals, say, desert ants, snakes, or cows, attribute mental states to others. Cows, presumably, don’t have mental states in their ontology. They see other cows as living bodies behaving in ways that make biological sense—eating, ruminating, sleeping, walking, and so on—rather than as agents carrying out decisions based on their desires and beliefs. For a few other particularly clever social species such as chimpanzees, dogs, crows, or dolphins, the question raised by Premack and Woodruff remains controversial: yes, these animals may well be capable of some rudimentary mindreading, but nothing approaching the virtuosity of humans in the matter.

  Premack and Woodruff’s article had a huge impact on the study not only of animal psychology but also of children’s. At what age do children start reading minds? A whole field of research initiated in the early 1980s showed that around the age of four, children readily attribute false beliefs to others (which, for good or bad reasons, has become the litmus test of genuine mindreading).6 And then, in 2005, a groundbreaking study by Onishi and Baillargeon,7 followed by many more studies confirming their findings, showed that not just four-year-olds but even infants are paying some attention to what others around them have in mind and even expect an agent’s actions to be consistent with its beliefs, whether true or false.

  Luca Surian, Stefana Caldi, and Dan Sperber, for instance, showed thirteen-month-old infants a video of a hungry caterpillar.8 The caterpillar, having seen a hand put a piece of cheese behind a screen on its left and an apple behind another screen on its right, went around the left screen and nibbled at the cheese (Figure 12). The infants saw this video several times so that it would be clear to them that the caterpillar had a preference for the cheese. Then came the crucial test phase. This time, infants saw the hand put the cheese behind the right screen and the apple behind the left screen. They saw the caterpillar arriving on the scene only after the two food items had been hidden, so that it wouldn’t know that their positions had been switched.

  Figure 12. In the familiarization phase, infants see a caterpillar go and nibble at a piece of cheese.

  What would, in such conditions, the infants expect the caterpillar to do? To go and look behind the left screen where it had repeatedly seen the cheese being hidden, or to go to the right screen behind which the cheese had, this time, been put? To us adults, it is obvious that the caterpillar would go where it had grounds to believe the cheese to be, that is, behind the left screen, even though this belief would, in this case, be false. But are infants able to draw this kind of inference and in particular to attribute false beliefs to others? Remarkably, infants behaved like adults would. They expected the caterpillar to go to the left, where it had grounds to believe (falsely) it would find the cheese. They looked longer when the caterpillar went straight to the cheese on the right side, where it had no way to know that the cheese was. In other words, infants expected an agent to behave according to its beliefs, whether true or false.

  Should we conclude that the infants in this study have a mental representation of a general psychological fact—that agents form beliefs and intentions rationally—and that the agents use this psychological fact as a premise in inference? Here again, it makes better sense to assume that what infants have is a specialized procedure that exploits some regularity in the way agents form beliefs and intentions. Infants need not represent this regularity.

  How can mindreading exploit the fact that agents such as humans and caterpillars tend to be rational without actually using a mental representation of this fact as a premise? What makes agents rational, we have suggested in Chapters 3 through 5, isn’t a general mechanism or disposition to think and act rationally, but a variety of inferential mechanisms with different inferential specializations. These mechanisms, notwithstanding their diversity, have all been shaped by a selective pressure for efficiency—in this case, for performing in an efficient way the kind of inferences it is their function to perform.

  Arguably, “rationality” in a most basic sense is synonymous with inferential efficiency. The degree to which rationality is achieved depends to a large extent on the way many inferential modules each perform their functions and on the way these modules are articulated. In order to take advantage of agents’ rationality, mindreading must then turn to specific inferential mechanisms and exploit their tendency to deliver, each in its domain, efficient inferences.

  To better understand how this multifaceted rationality of real people can be exploited in mindreading, we had better step out of the laboratory and away from the narrow focus on the “false belief task” and a few related experimental paradigms, however well designed they may be.

  Here is a simple example of everyday life mindreading. You enter the waiting room at your doctor’s office. There is already another patient. You both exchange glances and say, “Hello!” You sit down. She is intermittently typing on her smartphone and staring at the screen. You take a magazine. She looks at her watch and sighs. You exchange another glance. No doubt, you each have a train of personal thoughts that is opaque to the other, but still, you both do some light mutual mindreading as a matter of course.

  When you arrived in the waiting room, she knew that you would understand that she would see the doctor before you. You were disappointed that there was already someone waiting, but you tried not to show it, not to let her read your mind on this, but she probably did all the same. You understood that her alternately tapping on her smartphone and staring at it were part of an ongoing interaction with someone else with whom she was chatting at a distance (even if you had no idea what they were chatting about). You guessed that she was looking at her watch because it might already be past the time of her appointment with the doctor, and that she was sighing because she was unhappy to have to wait. You understood from the exchange of glances that she had understood that you had understood her sigh. And so on. All this mindreading that occurred in this minimal interaction was done spontaneously and without effort.

  The same kind of spontaneous mindreading occurs even in the absence of interaction, when one is just watching another agent that is unaware of being watched. The thirteen-month-old infants understand that the caterpillar, having seen the cheese repeatedly placed behind the left screen, assumes that it is again in the same place. The infants, who, unlike the caterpillar, have witnessed the cheese been placed on the right side this time nevertheless expect the caterpillar to search for the cheese on the left side, as before.

  Such mindreading is so obvious! What is not obvious, however, is what makes it so obvious. Here is a possible sketch of what happens. We humans tend to constantly mo
nitor our social environment (as in the waiting room example). We open, maintain, and update “mental files”9 on all the people we know (including people we only know of, like kings or famous actors, not to mention fictional characters whose thought we also know quite a bit about). In these files about people (and other kinds of agents such as caterpillars and gods), there may be all kinds of information: information about their names, family, history, appearance, dispositions, doings; information also about what is in their mental files where they store information about other people, us included. Our mental files about other people contain information about the contents of their mental files, and that information is provided by mindreading (and is, of course, metarepresentational).

  Some of your mental files about people are very thin and short-lived, such as the file you opened about the other patient in the doctor’s waiting room. Other files are thick and permanent, such as files about members of your close family. Some of the mindreading information in these files is provided by your spontaneous interpretation of what you observe others doing. Some information is provided by the people themselves who, in communicating, help you read their minds. Further information is provided through people talking about people: gossip. The point is that we read minds on the basis of a great variety of evidence and in a great variety of ways.

  There must be a mindreading module—actually a minds-reading module, with “minds” in the plural—that has the job of managing, in our mental files about other people, what these people have in their mental files. No such module, however, could do the job on its own. In order to perform mindreading inferences about the inferences that are performed in other people’s mind, the mindreading module must be linked to a great variety of other inferential modules and use them for updating the information represented in individual files.

  Take a very simple example. Tim asks you for candies. You give him some. He looks at them, says, “Three candies, eh?” and puts them in an empty paper bag. In the file you keep about Tim, there is now a metarepresentation of the belief you assume he has that there are three candies in the bag. He says, “Some more, please!” You give him another handful of candies, and he looks at them and says, “Five candies—thanks—with the first three, that’s enough,” and puts them in the bag. In your Tim file, there is now a metarepresentation of his belief that there are five more candies in his bag. As it happens, you counted not five but six more candies, so you believe that what Tim believes is wrong.

  On the basis of Tim’s beliefs that he first put in his bag three candies and that he then added five more, you might attribute to him the further (false) belief that there are eight candies in the bag. But how would you carry out this mindreading inference? Performing arithmetic operations—for instance the addition 3 + 5 = 8—is not something your mindreading module is equipped to do on its own. For this, your mindreading module has to share the content of the two relevant beliefs of Tim (that he first put three candies in the bag and that he then added five more) with your arithmetic module and copy back the output of its operation into Tim’s file, thus coming to metarepresent Tim’s belief that there are eight candies in the bag. Your arithmetic module, presented with your own belief that six candies were added to the initial three (rather than five, as Tim wrongly believes), drew the conclusion that there are now nine candies in the bag. This, however, goes in the mental file you have opened for the bag—we have mental files not just for people, but for all kinds of things—and not in the file you have opened for Tim.

  There is, actually, growing evidence that our highly social minds track and anticipate all the time not only what happens in our physical environment but also what happens in the minds of others around us.10 To achieve this double kind of tracking, many of our inferential modules routinely perform inferences not just to update our own beliefs about the world but also to update our metarepresentations of the beliefs of other people around us.

  Does this mean that we understand what others think by taking their point of view? Actually, only whole persons have a point of view and may attempt to see things from another person’s point of view. Mental modules in the individual’s brain, on the other hand, are “subpersonal”11 mechanisms and don’t have points of view.

  Is it that modules are occasionally used “offline” to simulate the mental processes of other people? We suggest, rather, that tracking the mental processes of others is part of their regular “online” job. Our arithmetic module, for instance, computes quantities in a perspective-neutral manner, and the output of these computations may update our representations of the way things are as well as our metarepresentations of the way things are represented in other people’s minds.12

  In the caterpillar experiment, infants see the cheese being repeatedly placed behind the left screen and see the caterpillar seeing the same thing. They draw, both for the caterpillar and for themselves, the inference that this is where the cheese regularly goes. When, in the absence of the caterpillar, the infants now see the cheese being placed behind the right screen, they update their own representation of the location of the cheese, but not their metarepresentation of the caterpillar’s representation. Later, when the caterpillar arrives on the scene, it is the belief metarepresented in its file (a belief that now is false) that the infants’ mindreading module uses to predict where it will look for the cheese.

  How, then, do infants form sensible expectations about the future actions of the caterpillar? We suggest that when the caterpillar arrives on the scene, the infants’ mindreading module (1) transfers information from the caterpillar’s file to a goal-directed-movement module, the job of which is to compute a rational path to a goal in space; and (2) uses the result of this computation to update the caterpillar’s file and to anticipate its movements.

  In the waiting-room example, you exploit, among other modules, a modularized competence informed by social conventions that guides you in the kind of situation where you happen to be for a while physically close to strangers because of goals that are parallel but not shared (as, for instance, in a waiting room, in an elevator, or on a plane). There, in the waiting room, you register the events relevant to your interaction simultaneously from your perspective and from that of the other patient. You interpret your brief salutations and your exchange of only short glances as a means to maintain the social distance you probably both feel most comfortable with. In order to do this mindreading, you do not have to actively decide to take the other person’s perspective. You were, from the start, automatically updating the file you opened about her mental states just as you were, in a much more fine-grained manner, updating your own files about the physical objects and events around you.

  This kind of automatic, permanent tracking of the mental states of others around us routinely involves, we are assuming, the attribution of beliefs, intentions, decisions, and other contentful mental states. It develops from infancy to adulthood (with individual and cultural variations) into a fairly elaborate capacity to understand others, so to speak, on the fly.13

  Cultural traditions diversely enhance, hinder, and otherwise influence the way we understand each other, and so do social roles and professional occupations. Across cultures and historical times and even within cultures, there is a great variety of explicit ideas about the human mind, which can be expressed in proverbs as well as in elaborate folk or scholarly theories. The mindreading we do every day in interacting with others (or in hearing or reading about them) remains, however, quite spontaneous and intuitive. It doesn’t use these cultural ideas as premises in spontaneous inference aimed at recognizing the mental states of others. These ideas, rather, are used, when need be, to help explain and justify conclusions that were arrived at quite intuitively.

  Virtual Domains

  Mindreading, which provides us with intuitions about people’s beliefs, desires, and intentions and about how these relate to what people perceive and do, by no means exhausts our ability to draw inferences about representations. Numerical cognitio
n, for instance, provides us with a sharply different kind of metarepresentational intuitions.

  As much recent work on numerical cognition has shown,14 human infants share with other animals the ability to mentally represent quantities of discrete objects. These representations are exact for very small quantities—one, two, and three—and are approximate for larger quantities. The acquisition of a language with names for numbers provides humans with lexical tools to represent in an exact manner quantities much larger than three. Linked to language and to writing, the cultural emergence of numeral systems, that is, of symbols used to produce public representations of quantities, has led in some cultures to the development of a whole new branch of knowledge, arithmetic. Anybody who has learned some arithmetic (and even, it seems, anybody who just uses numerals to count)15 has intuitions not only about concrete quantities but also about formal relationships among numbers apprehended through the numerical symbols that represent them.

  To give just one example, it is intuitively obvious to you—you don’t have to make any computation—that 900 is three times 300. This intuition exploits your knowledge of relations among very small numbers (in this case, that 3 × 3 = 9) and a property not of quantities but of a particular way of representing them, the decimal system (with Arabic numerals). If you used a base nine system, it wouldn’t be as immediately obvious that 1,210 is three times 363 (even though these numerals represent the same two numbers as do 900 and 300 in the decimal system). On the other hand, if you used a base nine system, it would be quite obvious to you that 1,000 is three times 300. It is with the decimal system this time that it is less immediately obvious that 729 is three times 243 (even though these numerals represent in base ten the same two numbers that 1,000 and 300 represent in base nine). We have better intuitions about rounded numbers; however, rounding isn’t a property of numbers but of the system of numerals used to represent them.