Game Theoretic Bias Detection in Multi-Agent AI Systems

TL;DR: Leverage game theory to uncover hidden biases in AI agents during strategic interactions, using frameworks like FAIRGAME to evaluate and mitigate fairness issues in real-world scenarios like collaborative decision-making or competitive environments.
Research Question: Can game-theoretic frameworks effectively detect and mitigate biases in multi-agent AI systems that emerge from interactive behaviors rather than just training data?
Hypothesis: Modeling agent interactions as strategic games will reveal emergent biases (e.g., in resource allocation or decision outcomes) that static evaluations miss, enabling targeted mitigations to improve overall system fairness and robustness across diverse agent profiles.
Experiment Plan: Adapt a game-theoretic framework (e.g., FAIRGAME) to a multi-agent LLM simulation, such as a prisoner's dilemma variant or auction scenario with agents varying in attributes like language or strategic traits. Inject potential bias triggers (e.g., uneven starting conditions) and run iterations to measure outcome disparities. Apply mitigation strategies like equilibrium adjustments, then compare pre- and post-mitigation on metrics including demographic parity, equalized odds, and task success rates. Analyze via case studies on bias propagation in group dynamics.

Reference:

FAIRGAME: a Framework for AI Agents Bias Recognition using Game Theory (arXiv, 2025): https://arxiv.org/abs/2504.14325 (full paper) or https://arxiv.org/html/2504.14325v3 (HTML version).
FairGamer: Evaluating Biases in the Application of Large Language Models to Video Games (arXiv, 2025): https://arxiv.org/abs/2508.17825 (full paper) or https://arxiv.org/pdf/2508.17825 (PDF version).
Game Theory Meets Large Language Models: A Systematic Survey (IJCAI, 2025): https://www.ijcai.org/proceedings/2025/1184 (conference proceedings) or https://arxiv.org/abs/2502.09053 (arXiv preprint).