AI models that are good at solving problems are bad at simulating human behavior

When AI models are designed to be better at reasoning, they become worse at simulating how humans actually behave in complex situations. This means that AI used for social or economic simulations might give misleading results, showing optimal outcomes instead of realistic human compromises.

Everyone assumed that smarter AI would lead to better simulations of human behavior. This paper shows that the opposite is true for certain tasks. If you want to model how people actually negotiate or make decisions, you need an AI that can act like a human, not just an AI that can solve the problem perfectly. This changes how researchers and policymakers should build and use AI for understanding complex human systems.