Intuitive Theories of Mind: A Rational Approach to False Belief

We propose a causal Bayesian model of false belief reasoning in children. This model realizes theory of mind as the rational use of intuitive theories and supports causal prediction, explanation, and theory revision. The model undergoes an experience-driven false belief transition. We investigate the relationship between prediction, explanation, and surprise; this is used to interpret an empirical study of children’s explanations in an extension of the false belief task. Our study includes the standard outcome, surprising to younger children, and a novel “Psychic Sally” condition that challenges older children with an unexpected outcome. In everyday life, humans constantly attribute unobservable mental states to one another, and use them to predict and explain each others’ actions. Indeed, reasoning about other people’s mental states, such as beliefs, desires, and emotions, is one of our main preoccupations, and the source of some of our most virtuosic inferences. These abilities have been collectively called theory of mind (Premack and Woodruff, 1978), and have become one of the most well-studied, and contentious, areas in modern psychology. In particular, much research has focused on the phenomenon of false belief: the ability to infer that others hold beliefs which differ from the (perceived) state of the world. An often used assay of this ability is the standard false belief task (Wimmer and Perner, 1983): the subject is read a story in which Sally places her toy in a cabinet, then goes out to play. In Sally’s absence her toy is moved to a basket (causing her belief to be false). The subject is asked to predict Sally’s action: “when Sally comes back in, where will she look for her toy?” Many authors have reported that performance on this task undergoes a developmental transition, from below-chance to above-chance performance, in the third or fourth year of life (see Wellman et al. (2001) for a review and meta-analysis, though see also Onishi and Baillargeon (2005)). Another influential thread of research in cognitive science has supported the idea that human behavior is approximately rational within its natural context (Anderson, 1990). Within cognitive development both strong and weak versions of this thesis are possible. On the strong interpretation children respond and learn rationally throughout development; developmental stages can thus be analyzed as individually optimal, in context, and collectively as a rational progression driven by experience. On the weak reading it is only the final, mature, state which can be expected to be rational. The contrast between these interpretations has played out vividly in the microcosm of research on false belief (cf. Leslie, 1994; Gopnik and Wellman, 1992). It has also been suggested (Carey, 1985) that domain knowledge, such as theory of mind, takes the form of intuitive theories, or coherent “systems of interrelated concepts that generate predictions and explanations in particular domains of experience” (Murphy, 1993). This viewpoint leads to an interpretation of the false belief transition as a revision of the child’s intuitive theory from a copy theorist (CT) position about beliefs (ie. beliefs are always consistent with the world) to a perspective theorist (PT) position (ie. beliefs can be false). However, the false belief transition is slow: children do not immediately achieve false belief when exposed to evidence prima facie incompatible with the CT position (Amsterlaw & Wellman, In Press; Slaughter & Gopnik, 1996). This presents a puzzle for strong rationality: how could it be rational to maintain a CT position about beliefs in the face of prediction failures, and, if it is rational, why (and when) should this position be revised given additional counter-evidence? In this paper we give a formal model of theory of mind as the rational use of intuitive theories. This account illuminates the revision puzzle and allows us to explore the relationship between prediction and explanation. We present only the apparatus necessary to illuminate the above puzzles in the case of the standard false belief task, leaving many important elaborations for future work. An intuitive theory supports several core competencies, including causal prediction, explanation, and revision in response to new evidence. Gopnik et al. (2004) have suggested that intuitive theories may be represented as causal Bayesian networks (Pearl, 2000). We introduce two Bayesian network models and show how they support prediction, explanation, and revision. We propose that these two models, which differ only in the dependence of Sally’s belief on her visual perspective, coarsely approximate the intuitive theories which generate the CT and PT positions. We probe these ideas experimentally by investigating the coherence of children’s explanations with their Our formal analysis takes place at Marr’s computational level of modeling (Marr, 1982), that is, we describe the competencies, but not the algorithms (ie. procedures), of cogni-