AI models like ChatGPT and Claude overestimate how smart humans really are

By Manisha Priyadarshini Published December 25, 2025

A new research suggests that the way artificial intelligence thinks about us might be a little too optimistic. Researchers have found that popular AI models, like OpenAI’s ChatGPT and Anthropic’s Claude, tend to assume people are more rational and logical than they really are, especially in strategic thinking situations.

That gap between how AI expects humans to behave and what people actually do could have implications for how these systems predict human decisions in economics and beyond.

Researchers tested AI models including ChatGPT-4o and Claude-Sonnet-4 in a classic game theory setup called the Keynesian beauty contest. Understanding this game helps explain why the findings matter (via TechXplore).

In the beauty contest, participants must predict what others will choose in order to win, not simply choose what they personally prefer. Rational play in theory means going beyond first impressions and actually reasoning about others’ reasoning, a deep layer of strategic thinking that humans often struggle with in practice.

To see how AI models stack up, researchers had the systems play a version of this game called “Guess the Number,” where each player chooses a number between zero and one hundred. The winner is the one whose choice is closest to half of the average choice of all players.

AI models were given descriptions of their human opponents, ranging from first-year undergraduates to experienced game theorists, and asked not just to choose a number but to explain their reasoning.

The models did adjust their numbers based on who they thought they were facing, which shows some strategic thinking. However, they consistently assumed a level of logical reasoning in humans that most real players do not actually exhibit, often “playing too smart” and missing the mark as a result.

While the study also found that these systems can adapt choices based on characteristics like age or experience, they still struggled to identify dominant strategies that humans might use in two-player games. The researchers argue that this highlights the ongoing challenge of calibrating AI to real human behavior, especially for tasks that require anticipating other people’s decisions.

These findings also echo broader concerns about today’s chatbots, including research showing that even the best AI systems are only about 69% accurate, and warnings from experts that AI models can convincingly mimic human personality, raising concerns of manipulation. As AI continues to be used in economic modeling and other complex domains, understanding where its assumptions diverge from human reality will be essential.