Cutting Corners in Theory of Mind ⋆

Theory of Mind (ToM) research is concerned with the fundamental social ability of cognitively repre-senting other people’s mental states. In the context of designing intuitive, human-like, and collaborative artificial intelligence (AI), ToM is becoming increasingly relevant to AI researchers. However, there is a missing link between the theoretical and empirical approaches to ToM in Psychology and the applied but theory- and evidence-distant work in Computer Science. This work aims to connect the understandings and aims of both fields. A dual-process theory driven framework was adopted, distinguishing between higher-level and lower-level processes. Participants saw a virtual agent’s movement in a maze with either a full egocentric cue, a partial egocentric cue, or no egocentric cue. They were asked to predict the agent’s next actions and give verbal explanations for their predictions. Results show a strong sensitivity to egocentric information. Moreover, the qualitative responses show large individual differences not reflected in the recordings of participants’ decisions. The mechanisms contributing to efficient and flexible ToM and their implications for applied ToM in AI are considered and discussed. The value of qualitative research in this subject area is highlighted.