Large language models replicate and predict human cooperation across experiments in game theory

What Happened

A new arXiv paper (2511.04500v3) presents a systematic investigation into how well large language models replicate and predict human cooperative behavior across classic game theory experiments. The researchers tested multiple LLMs on established experimental paradigms—including Prisoner’s Dilemma, Public Goods games, and Trust games—comparing model outputs against known human behavioral data. The core finding is that modern LLMs, particularly GPT-4 and Claude 3, demonstrate surprisingly high fidelity in mimicking human cooperation patterns, including nuanced behaviors like conditional cooperation, reciprocity, and sensitivity to framing effects. The models not only reproduced aggregate human choices but also captured individual-level variation in strategic decision-making.

Why It Matters

This research carries significant weight for two distinct audiences. First, for behavioral scientists, it suggests LLMs could serve as cost-effective, scalable proxies for human subjects in pilot studies or exploratory research. Rather than running expensive human experiments to test new game theory hypotheses, researchers might first probe LLMs to identify promising directions. However, the paper also warns that LLMs are not perfect mirrors—they tend to over-cooperate in some settings and under-react to certain social cues, meaning they cannot replace human validation.

Second, for AI deployment in high-stakes domains—such as automated negotiation, resource allocation, or collaborative AI systems—the findings have practical implications. If LLMs can accurately predict human cooperative strategies, they could be used to design better human-AI interaction protocols. Conversely, the fact that LLMs exhibit human-like biases in cooperation (e.g., ingroup favoritism, punishment of defectors) means developers must be cautious about deploying these models in settings where impartial decision-making is critical. An AI that “cooperates” too readily might be exploited, while one that defects too frequently could break trust.

Implications for AI Practitioners

For those building or deploying LLM-based agents, this research offers both reassurance and a warning. The positive takeaway is that LLMs can serve as reasonable behavioral simulators for testing multi-agent systems or designing cooperative AI. Practitioners could use LLM-generated cooperation data to train reinforcement learning agents or to prototype negotiation strategies before real-world deployment.

The cautionary note is that LLM cooperation patterns are not static—they vary by model, prompting strategy, and even by the specific framing of the game. This means developers cannot assume a single “cooperation personality” for an LLM agent. Instead, they must systematically test how their model behaves across different game-theoretic contexts, especially when the stakes involve real resources or human trust. Additionally, the paper highlights that LLMs trained on human text may inherit both the best and worst of human social behavior—including tendencies toward conditional cooperation that can lead to cascading defection in multi-agent settings.

Key Takeaways

• LLMs can replicate human cooperation patterns across multiple game theory experiments, making them useful as behavioral proxies for exploratory research and simulation.
• Models are not perfect substitutes for human subjects—they exhibit systematic biases, including over-cooperation and reduced sensitivity to certain social context cues.
• Developers must test LLM agents contextually, as cooperation behavior varies by model, prompting, and game framing, with no single “default” cooperative strategy.
• Human-like biases in LLM cooperation (e.g., reciprocity, ingroup favoritism) pose risks for high-stakes deployment, requiring careful design of guardrails and evaluation protocols.