fMRI validation of GPT-4's ability to recognise Theory of Mind in natural conversations

Overview & Synopsis

Several studies in Artificial Intelligence emphasise the potential of Large Language Models (LLMs) in spontaneously developing Theory of Mind (ToM) due to their exposure to human-like language expressing mental states and reading literary fiction. Recent evaluations of GPT-4's ToM capabilities provide evidence that GPT-4 demonstrates advanced reasoning about the mental states of characters, adept handling of abstract situations, and proposing cooperative actions in social contexts. Here, we introduce the validation of ToM tagging, a tool leveraging LLMs to identify ToM exchanges within transcripts of naturalistic social interactions, using fMRI human brain recordings.