Theory of Mind in Large Language Models: Can AI Understand Human Thoughts?

This video reviews studies discussing whether large language models (LLMs) such as GPT-4 have theory of mind (ToM). Theory of mind is the ability to understand other people's thoughts, beliefs, and intentions — a fundamental aspect of human social cognition.

In his article (Kosinski, 2024), Michal Kosinski presents a study showing how later LLMs (like ChatGPT-4) can successfully solve complicated false belief testing tasks. He argues that this demonstrates these LLMs might possess understanding comparable to six-year-old children, potentially emerging from the way language skills develop during training.

However, a critical article by Hu, Sosa, and Ullman (Hu et al., 2025) challenges current evaluation methods. The authors argue that LLM success in ToM tasks often reflects only a coincidence of behavioural outcomes with humans, rather than the use of similar computational processes. This distinction is a key reason for conflicting conclusions in this field.

The critics propose shifting focus from whether AI systems behave similarly to humans toward understanding how LLMs actually work — ensuring a more rigorous scientific understanding of machine cognition.

Evaluating large language models in theory of mind tasks

Kosinski, Michal. Proceedings of the National Academy of Sciences 121, no. 45 (November 2024)

Re-evaluating Theory of Mind evaluation in large language models

Hu, Jennifer, Felix Sosa, and Tomer Ullman. Preprint (2025)