When you sit down at a game where choices matter and money is on the line, do you expect your artificial opponent to think like you — or to surprise you with unexpected cooperation? A new laboratory study offers a clear, if unsettling, answer: people change how they play when they know an opponent is a large language model (LLM).
What the experiment did and found
The study reported the results of the first controlled, monetarily-incentivised laboratory experiment to compare human behaviour in multi‑player strategic settings when opponents are other humans versus when opponents are LLMs. Researchers used a within‑subject design to track how the same individuals behaved across conditions, and they implemented a multi‑player p‑beauty contest — a simultaneous choice game that emphasizes strategic reasoning.
The core empirical result is unambiguous: human subjects chose significantly lower numbers when they believed they were playing against LLMs than when they were playing against humans. The change in play manifested largely as an increased prevalence of choices at the 'zero' Nash‑equilibrium.
Who drove the change — and why
The shift toward lower choices was concentrated among subjects with higher strategic reasoning ability. In other words, players who are better at anticipating others' decisions were more likely to adopt the lower, equilibrium strategy when facing LLM opponents.
Participants who selected the 'zero' Nash‑equilibrium explicitly motivated their strategy by appealing to two perceptions about the LLM opponents: first, the LLMs' reasoning ability; and second, an unexpected belief in the LLMs' propensity toward cooperation. That combination — expecting both stronger strategic thinking and a willingness to cooperate — appears to have pushed well‑reasoning players toward the pure equilibrium choice.
What this suggests for mixed human–LLM systems
The authors frame these findings as foundational for understanding multi‑player human–LLM interaction in simultaneous choice games. The experiment uncovers heterogeneities both in subjects' behaviour and in their beliefs about how LLMs will play when mixed into human strategic environments.
Crucially, the study concludes by flagging design consequences: these behavioural shifts suggest important implications for mechanism design in systems that combine humans and LLM agents. If human players systematically alter their strategies based on expectations of LLM reasoning and cooperation, then designers of markets, platforms, and decision protocols that mix humans and LLMs will need to account for those expectations when crafting incentives and rules.
How to read the results — perspectives and puzzles
- Technologists: The findings imply that an LLM's apparent reasoning competence and perceived cooperativeness can change human strategic choices. That matters for designers who embed LLMs in negotiation, coordination, or competitive environments.
- Policymakers and mechanism designers: The study directly links human beliefs about LLMs to measurable shifts in behavior, which raises questions about how institutions should regulate or structure interactions that include artificial agents.
- Users: Even without explicit coordination, some humans expect LLM opponents to be both rational and cooperative; that expectation can lead to markedly different choices than when humans face other humans.
- Researchers and adversaries: The heterogeneity of responses — especially concentrated among high‑reasoning subjects — points to nuanced effects that could be exploited or mitigated, depending on the application and incentives at play.
The research does not claim to answer every question about human–LLM strategic interaction, but it does establish a clear pattern: knowing an opponent is an LLM changes what people do, and those changes are driven by beliefs about the LLMs' reasoning and cooperative tendencies. That pattern is small in description but large in consequence — because strategic environments cascade: a single shift toward the Nash equilibrium in a multi‑player game can alter outcomes for many participants.
If humans expect both greater rationality and surprising cooperation from LLMs, how should we design systems where those expectations matter — and who benefits when they are wrong? The study opens the door to those questions and makes one thing plain: mixed human–LLM systems will not behave like purely human systems, and designers who ignore that risk will find their mechanisms produce different incentives and outcomes than intended.
https://www.schneier.com/blog/archives/2026/04/human-trust-of-ai-agents.html
