Emergent Social Bias in AI and Mitigation Strategy

Bias in artificial intelligence is commonly viewed as a reflection of human and societal prejudices embedded in training data. This study instead examines whether bias can emerge endogenously from interaction and learning among AI agents, even in the absence of human-generated data or societal input. Using deep reinforcement learning (DRL) agents trained solely through repeated interactions with one another in a trust game, we find that the agents autonomously develop in-group favoritism and out-group discrimination, which reduces overall trust and trustworthiness at the population level. We evaluate multiple mitigation strategies and find that training-based interventions, such as inclusive training and retraining with unbiased AI coaches, reduce group disparities but fail to improve, and often worsen, cooperation among AI agents. In contrast, an architectural intervention that disrupts internal neural representations can eliminate group biases while maintaining high levels of cooperation among agents. To understand the origin of emergent bias in AI, we introduce an information-theoretic framework based on mutual information. Our analysis suggests that group identity information and behaviorally relevant information may be encoded differently in DRL agents' neural networks. These findings motivate new design principles for building socially intelligent and cooperative multi-agent AI systems.