A new study from researchers at the University of Birmingham and other institutions shows that large language models (LLMs) can use mentalization—the ability to infer others' beliefs and intentions to guide decisions—in ways that sometimes surpass human performance. In two economic games, GPT-5 demonstrated flexible, adaptive reasoning that outperformed 251 human participants, while other models varied widely.
The Research
Scientists led by Lei Zhang and colleagues tested 2,099 individual LLM agents from four model families—DeepSeek, GPT-4.1, GPT-5, and Gemini 2.0 Flash—against opponents of varying sophistication. They used two economic games that require understanding what another player believes, using cognitive computational modeling to uncover latent strategies. The study, published on arXiv in August 2026, also benchmarked the AI agents against 251 human participants.
Results showed that LLMs displayed clear behavioral and computational signatures of mentalization, but their abilities differed markedly by provider and model size. Strategic prompting, which encouraged deeper reasoning, generally improved performance, though the benefit varied by task. Notably, GPT-5 agents adapted their recursive depth of reasoning to increasingly sophisticated opponents, demonstrating superior performance to humans. Other models, like DeepSeek, showed less flexibility.
Why It Matters
Mentalization is central to human social interactions—it helps you predict others' actions and make smart decisions. This research provides a formal method to assess this ability in machines, using cognitive computational models that are also used to study human cognition. For you, understanding that AI can engage in mentalization offers a way to reflect on your own strategic thinking. It also highlights that AI models differ in their social reasoning, which has implications for how we use them as assistants or partners.
What You Can Do
To sharpen your own mentalizing skills, practice perspective-taking in everyday situations. When bargaining or collaborating, consciously ask: What does this person believe about me? What are their goals? You can also play strategic games like chess or poker that require anticipating others' moves. For a more formal assessment, our free IQ test includes puzzles that tap into logical and social reasoning.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.