How Humans Behave Against LLMs in Strategic Games: Rationality and Cooperation

Research into how humans interact with artificial intelligence in strategic decision-making has revealed a notable pattern: people tend to expect AI agents to act more rationally and cooperatively than human counterparts. This insight comes from a controlled laboratory experiment published in May 2025, which examined behavior in a multi-player p-beauty contest game when participants believed they were competing against either humans or large language models (LLMs).

The study, conducted by researchers Darija Barak and Miguel Costa-Gomes, found that when individuals thought they were playing against an LLM, they chose significantly lower numbers compared to when they believed their opponents were human. This shift was primarily driven by an increase in selections of zero—the Nash equilibrium in the game—suggesting participants anticipated the AI would reason to the logical conclusion of the game more consistently than humans might.

Further analysis indicated that this behavior was most pronounced among participants with higher strategic reasoning ability. These individuals explicitly justified their choice of zero by referencing both the perceived superior reasoning capacity of LLMs and, unexpectedly, an assumption that the AI would be more inclined toward cooperation. The researchers noted this combination of expectations—high rationality paired with presumed cooperativeness—was not initially anticipated and highlights nuanced beliefs humans hold about AI behavior in interactive settings.

The p-beauty contest game used in the experiment is a classic tool in behavioral economics for studying iterative reasoning. In this game, players select a number between 0 and 100, with the winner being whoever chooses closest to a fraction (typically two-thirds) of the average of all guesses. The Nash equilibrium occurs when all players select zero, as any deviation would not be optimal if others are also playing rationally. Reaching this equilibrium requires multiple steps of iterative thinking: anticipating others’ choices, then anticipating their anticipations and so on.

By comparing choices in human-vs-human and human-vs-LLM conditions within the same subjects, the researchers isolated the effect of opponent type on decision-making. The within-subject design strengthened the validity of the findings by controlling for individual differences in risk tolerance, strategic sophistication, and other personal variables. The monetarily incentivized nature of the experiment ensured participants had a real stake in the outcome, increasing the ecological validity of the observed behaviors.

The results suggest that when humans interact with AI systems in strategic contexts—such as negotiations, market competitions, or collaborative problem-solving—they may bring assumptions about the AI’s cognitive style that significantly influence their own behavior. These assumptions are not merely about computational power but extend to perceived traits like cooperativeness, which may not align with the AI’s actual programming or incentives.

Such findings have implications for the design of AI systems intended to operate in mixed human-AI environments. If users consistently expect LLMs to cooperate or act in ways that are not strategically optimal for the AI, mechanism designers may demand to account for these beliefs to predict interaction outcomes accurately. Conversely, if AI systems are designed to exploit these expectations—by appearing cooperative while pursuing self-interested strategies—it could raise ethical concerns about manipulation and trust erosion.

The study’s authors emphasize that this work provides foundational insights into the heterogeneity of human beliefs about AI. Not all participants reacted the same way. variations in strategic ability led to divergent responses, suggesting that tailored approaches may be necessary when deploying AI in diverse user populations. Future research could explore how these expectations evolve with repeated interaction, transparency about AI design, or exposure to AI behavior that contradicts initial assumptions.

As LLMs become more integrated into everyday economic and social interactions—from customer service bots to algorithmic trading agents—understanding the psychological dynamics of human-AI engagement becomes increasingly critical. This research contributes to a growing body of work examining trust, attribution of mental states, and behavioral adaptation in hybrid human-AI systems.

For readers interested in the methodology and full results, the original paper is available through the arXiv preprint server under the identifier arXiv:2505.11011. The study represents one of the first incentivized laboratory experiments directly comparing human behavior toward human and LLM opponents in a strategic game setting.

Moving forward, scholars and practitioners in AI ethics, behavioral economics, and human-computer interaction will likely build on these findings to develop better models of trust and expectation in AI-mediated interactions. As AI systems take on more autonomous roles in society, aligning technical design with accurate models of human cognition—not just assuming uniformity—will be essential for creating systems that are both effective and socially beneficial.

World Today Journal will continue to monitor developments in this space, particularly studies that examine how trust in AI forms, changes, and influences behavior across different contexts, and populations.

We invite our global audience to share their experiences and perspectives on interacting with AI in decision-making scenarios. Have you noticed yourself adjusting your strategy when you suspect you’re competing against an AI? What assumptions do you build about how it thinks or what it values? Join the conversation in the comments below.