Emergent Group Bias in AI Agents and Mitigation Strategies, a seminar by Professor Yan (Diana) Wu (San José State University)

Emergent Group Bias in AI Agents and Mitigation Strategies

By Professor Yan (Diana) Wu from San José State University

Abstract

Bias in artificial intelligence is often attributed to human and societal biases embedded in training data. We examine whether group-conditioned disparities can instead emerge endogenously through interaction and learning among AI agents. Using deep reinforcement learning agents trained through repeated interactions in a trust game, we find persistent differences in how agents treat in-group and out-group partners. Our analyses suggest that interaction structure is an important driver: group disparities become stronger when agents interact disproportionately with members of their own group, while more balanced exposure substantially changes the learned behavior. We also evaluate mitigation strategies, including inclusive training, AI coaching, and neural-network interventions. These approaches can reduce group disparities, although some involve tradeoffs with overall cooperation. The findings point to repeated interaction, co-adaptation, and learned internal representations as important factors in the emergence and mitigation of group bias in multi-agent AI systems.

About the speaker

Dr. Yan (Diana) Wu is Professor and Chair of Operations and Supply Chain Management at San José State University’s Lucas College and Graduate School of Business. She earned her Ph.D. in Business Administration and Operations Research from Penn State University. Her research focuses on operations management, behavioral economics, and artificial intelligence, with publications in Management Science, Production and Operations Management, and the Journal of Operations Management. She has also served as a visiting scholar at Yahoo! Labs and Hewlett-Packard Labs.