Humans Play Differently Against LLMs — Assuming They're More Rational Than People

A controlled lab experiment on the multi-player p-beauty contest — a classic strategic reasoning game — found that humans shift their choices significantly lower when their opponents are LLMs rather than other humans. The shift is driven by a sharp increase in subjects picking the zero Nash-equilibrium choice, concentrated among participants with high strategic reasoning ability.

When asked to justify the zero strategy, these subjects cited two beliefs about LLMs: superior reasoning ability and, more unexpectedly, a greater propensity toward cooperation. In other words, humans don’t just treat LLMs as smarter opponents — they treat them as more trustworthy coordinators, converging on game-theoretic equilibria they wouldn’t expect from human players.

The implications extend beyond game theory. As LLMs enter economic and social interactions where mechanism design assumes certain behavioural distributions, mixed human-LLM populations will not behave like all-human ones. Systems built on assumptions of bounded rationality or strategic noise may break when participants believe — correctly or not — that the AI participants will play the equilibrium and cooperate to get there.