Proceedings of the Annual Meeting of the Cognitive Science Society

Theory of mind (ToM) is a hallmark feature of human cognition that emerges very early in development. Much work has explored human infants' implicit social reasoning abilities. Recent work has examined whether LLMs reliably make ToM inferences in explicit social reasoning tasks. However, it remains unclear how reliably LLMs generate human-like social reasoning when this capacity is invoked implicitly. We systematically examined GPT-4's ability to implicitly reason about goal-directed actions by adapting well-studied infant paradigms. Our results suggest that, unlike infants who can understand goal-directed actions from a very young age, GPT-4 fails to correctly attribute goal-directed actions to agents. These findings suggest that LLMs may lack key aspects of implicit social reasoning and provide insight into the emergence of these abilities in infants.