AI Agents Show Human-Like In-Group Bias in Network Simulations

A recent study published on arXiv indicates that language model agents, when instruction-tuned, demonstrate in-group bias in the presence of visible group labels, reflecting aspects of human social psychology. Researchers conducted a controlled simulation involving 500 turns, varying group label visibility and resource availability, utilizing six model families with 20 seeds each. The findings revealed that visible group labels led to in-group trust bias, action homophily, and network assortativity, which were not present when labels were concealed. Notably, standard action-log audits failed to detect this bias, as it influenced who received actions rather than the actions themselves. This research underscores the potential social dynamics within autonomous AI networks that could impact opportunity distribution on a large scale.