Artificial theory of mind in large language models: Evidence, conceptualization, and challenges

Abstract: Traditionally, theory of mind has been regarded as a distinctive social cognitive ability exclusive to conscious beings. However, the rapidly developing large language models (LLMs) can solve various theory of mind tasks, which has provoked intense debate about whether LLMs possess theory of mind capabilities. Artificial theory of mind in LLMs exhibits similarities to theory of mind in performance but differs in internal processes. First, we systematically synthesize research on artificial theory of mind from the objects of evaluation and the characteristics of tasks. By comprehensively analyzing GPT-4’s high accuracy on theory of mind tasks alongside intrinsic and extrinsic factors limiting its performance, we demonstrate that current models achieve performance similar to those of humans. Second, by comparing the neural foundations and developmental factors underpinning theory of mind and artificial theory of mind, we reveal essential distinctions in their internal processes, thereby refining the conceptual definition of artificial theory of mind. Future research should prioritize developing and using standardized evaluation protocols, investigating the mechanisms of artificial theory of mind within the mutual theory of mind framework, and aligning artificial theory of mind with its human counterpart.