Discovering Differences in Strategic Behavior Between Humans and LLMs
By: Caroline Wang, Daniel Kasenberg, Kim Stachenfeld, Pablo Samuel Castro
Published: 2026-02-10
View on arXiv →Abstract
As Large Language Models (LLMs) are increasingly deployed in social and strategic scenarios, it becomes critical to understand where and why their behavior diverges from that of humans. While behavioral game theory (BGT) provides a framework for analyzing behavior, existing models do not fully capture the idiosyncratic behavior of humans or black-box, non-human agents like LLMs. This paper provides a comprehensive empirical analysis of LLM behavior in various strategic games, identifying systematic deviations from human decision-making patterns and offering insights into the implications for designing robust and ethically aligned AI systems in multi-agent environments.
Impact
practical
Topics
6
💡 Simple Explanation
Scientists made AI chatbots play psychological games involving money and strategy against each other and compared the results to how humans play. They found that AIs are generally too nice, cooperative, and afraid of risk, likely because they were trained to be polite assistants. This means AIs might currently be bad at tough business negotiations but great at cooperation.
🎯 Problem Statement
As LLMs are increasingly deployed as autonomous agents, it is unknown whether their strategic behavior aligns with human norms or mathematical rationality, creating risks in automated negotiations and economic modeling.
🔬 Methodology
Comparative analysis using a suite of 10 standard economic games (e.g., Prisoner's Dilemma, Stag Hunt). Models (GPT-4, Llama-3, Claude) were prompted with rules and asked to make decisions. Results were compared against historical human data from experimental economics literature using statistical distance metrics.
📊 Results
LLMs demonstrated a strong bias towards cooperation (70-90% cooperation rates in Prisoner's Dilemma) compared to humans (40-60%). They often failed to exploit weaker opponents, preferring 'fair' outcomes even when explicitly instructed to maximize profit. However, Chain-of-Thought prompting moved their behavior closer to Nash Equilibrium.
✨ Key Takeaways
Current LLMs are 'Good Samaritans' by default due to RLHF, making them poor proxies for competitive human behavior without specific fine-tuning. For realistic simulations, we must account for this 'niceness tax' or develop specific 'selfish' models.
🔍 Critical Analysis
The paper provides a compelling snapshot of current model limitations but potentially conflates 'training alignment' with 'intrinsic inability.' It is possible that prompt engineering was insufficient to unlock 'ruthless' behaviors. Furthermore, comparing text-based simulation to real-stakes human behavior always carries an ecological validity risk.
💰 Practical Applications
- Consulting services for optimizing AI agent parameters for specific market conditions.
- Licensing 'Human-Aligned' datasets for RLHF to make models less robotic.
- SaaS platform for automated dispute resolution using fair-play AI.