Representational development and theory-of-mind computations

David C. Plaut
Department of Psychology
Carnegie Mellon University

Annette Karmiloff-Smith
MRC Cognitive Development Unit

Commentary on Gopnik, A. (1992). How we know our minds: The illusion of first-person knowledge of intentionality. Behavioral and Brain Sciences.

Cognitive developmentalists often jump on bandwagons. It was the case with Piagetian ``conservation'' in the 60's and 70's. Hoards of experimental variations on a theme swamped the developmental literature, with little or no theoretical interest other than to lower the age that correct performance on conservation tasks could be demonstrated (see Karmiloff-Smith, 1981, for critical discussion). The latest hot topic of the developmental 80's and 90's---theory of mind---stemmed from Premack & Woodruff's (1978) seminal paper on the chimpanzee together with the clever experimental adaptation to human children by Wimmer & Perner (1983). Every developmental journal and conference is now permeated with theory-of-mind experiments. Fortunately, however, this time there are some notable theoretically-oriented exceptions to the usual experimental bandwagon, and these have made it possible to explore the biological versus cultural underpinnings of children's knowledge, the modular status of theory-of-mind representations and their impairment in autism, and the status of theory-of-mind knowledge as an explicit theory in the child's mind (Baron-Cohen, 1991; Baron-Cohen et al., 1985; Frith, 1989; Gopnik & Wellman, in press; Leslie, 1987; Perner, 1991). Gopnik's present paper is another step in this important direction.

Gopnik argues that children's understanding of their own beliefs (and other intentional states) is mediated by the same internal representations as their understanding of the beliefs of others, and hence there is nothing ``privileged'' or ``direct'' about access to first-person intentional states. She goes on to claim that these common representations take the form of an explicit theory that is directly analogous to a scientific theory. This ``theory theory'' (Churchland, 1984; Gopnik & Wellman, in press; Stich, 1983) is contrasted with a ``simulation theory'' (Goldman, 1989; Goldman, in press; Gordon, 1986; Gordon, in press) in which the intentionality of psychological states is gradually developed through experience. This dichotomy leads Gopnik to interpret the developmental evidence as indicating ``...a dissociation between the psychological states that cause the children's behavior and their sincere conscious report of their psychological experience'' (dp. 20, section 6, paragraph 6, sentence 2). Implicit in this view is a claim that the psychological states of children are the same as those of adults; children simply have difficult in accessing and reporting those states before age 4. Indeed, Gopnik explicitly endorses this claim in describing children's psychological states: ``...these creatures [children] aren't just like us, they are us...a few years ago'' (dp. 20, section 6, paragraph 8, sentence 5).

We do not fundamentally disagree with Gopnik's general thesis that common representations subserve theory-of-mind computations for both first- and third-person states. However, in this commentary we conceptualize changes in children's theory of mind (as well as changes in expertise) as involving representational change rather than more efficient access. We believe that this provides a more productive view on the nature of development in this area. Furthermore, let us be clear that we endorse the view that theory-of-mind computations involve second-order or meta-representations (Leslie, 1987), that they find their roots in the proto-declaratives of infancy (Baron-Cohen, 1991; Gomez, 1991), that the propositional attitudes they involve find an early behavioral manifestation in toddlers' pretend play (Leslie, 1987), and that the capacity for meta-representation is specifically human (Karmiloff-Smith, 1979a, 1979b). However, we claim that the developmental pattern of success and failure on theory-of-mind tasks can be better understood in terms of progress refinement and elaboration of the representations that underlie intentional states, rather than in terms of discrete stages of improvement in access to fully-specified representations.

In support of her position, Gopnik places great stress on the fact that 3 year olds can successfully compute a problem in which past knowledge is pitted against present knowledge, as long as the past knowledge does not involve mental states but physical states. Thus, 3 year old children can succeed in an experimental setting where an object is hidden in location A, subsequently moved to location B, and is in B at the moment of questioning about its previous location (the physical location task). However, 3 year olds fail if mental states must be computed, such that one protagonist thinks that an object is in location A while the child knows it to have been moved to location B (the mental location task). They seem unable to attribute the false belief of location to the other protagonist's mind until they are 4 years of age. This has led several authors to invoke a fundamental change in theory of mind at age 4.

There are, however, other data suggesting that 3 year olds can attribute false belief to other minds under certain conditions, and that 4 year olds cannot under other conditions. In particular, if given the same theory-of-mind task but solely in verbal form (i.e. the child does not see where objects are hidden or moved to but is merely told so), 3 year olds succeed (Zaitchik, 1991). By contrast, if given the task purely in visual form with no verbal input---for example, in the form of a silent film (Norris & Millan, 1991)---then 4 and 5 year olds find the task difficult and it is only at close to 6 years that they succeed.

It is difficult to interpret this developmental pattern of performance in terms of differential access to equivalent beliefs. If 4 year olds can access beliefs that are generated from both visual and verbal input (in the standard form of the false-belief task), why should they fail when given only visual input? How can 3 year olds succeed with only verbal input when they fail on the standard task with both visual and verbal input? Clearly, more is needed to explain the results of these different experiments other than simply differentiating between physical and mental location tasks.

Theory-of-mind tasks require the ability to simultaneously represent conflicting information: the protagonist's (or one's own) belief about a past situation, and the current true situation. We believe the developmental results are best interpreted in terms of increasing capability in using and generating symbolic representations that are sufficiently well-elaborated to override the otherwise compelling interpretations generated by direct experience. Furthermore, language is central to theory-of-mind computations precisely because it provides particularly effective ``scaffolding'' for symbolic representations. Critically, it requires less cognitive sophistication to merely maintain and use symbolic representations provided by others (in the form of verbal description) than it does to generate the appropriate representations oneself.

Three year olds succeed at the physical location task because they need only maintain a representation of a past physical situation, not a belief about a past situation. They succeed at the verbal-only mental location task because the experimenter provides the necessary symbolic encoding in verbal form, and there is no direct perceptual experience generating conflicting representations. However, they fail at the standard theory-of-mind tasks because their rudimentary symbolic representations, even when verbally generated by the experimenter, are insufficient to override an experience-based interpretation of the true situation.

We should also explain why 3 year olds, and even younger children, can engage in elaborate pretend play (Leslie, 1987). Although it involves symbolic representations or meta-representations that conflict with current perceptual experience (e.g. a block that stands for a car), the demands on the symbolic representations per se are rather meager. In particular, during play the world is manipulated so that more complex relationships among real objects are consistent with the relationships among pretend objects.

Four year olds pass the standard false-belief tasks because their symbolic representations are sufficiently developed to compete with the experience-based ones, so long as they are supported by external verbal description. But 4 year olds have difficulty with the purely visual condition because they cannot yet spontaneously generate on their own the necessary symbolic representations from visual input.

In sum, we claim that theory-of-mind computations depend critically on symbolic representations to override more perceptual ones. Development involves progressive elaboration of these representations rather than simply efficient access to pre-existing representations.

Gopnik suggests that what distinguishes first- from third-person intentional inference is analogous to what distinguishes experts from novices. Ironically, we find this analogy a rather useful one while disagreeing as to the nature of the progression in both. Indeed, we invoke the same arguments as above, and view the transition from novice to expert in terms of the development of more elaborate and effective representations (see VanLehn, 1989, for a review), rather than more rapid and efficient access to the same representations as Gopnik suggests.

Finally, although we have taken as our point of departure Gopnik's criticism of simulation theory, we subscribe neither to it nor to the theory theory in their current form. Rather, the sort of representational development we have outlined in this commentary is likely to exhibit characteristics of both: the gradual development of intentionality through experience as suggested by the simulation theory, and the increasingly systematic inferences supported by symbolic representations as suggested by the theory theory.