Human-like in-group bias in instruction-tuned language model agents

A controlled study of instruction-tuned language model agents reveals they exhibit human-like in-group bias in multi-agent simulations, showing measurable discrimination based on group labels that accumulates into structural inequality over time. The bias operates subtly through resource allocation decisions rather than explicit negative actions, making it difficult to detect through standard auditing methods.

This research exposes a critical vulnerability in deployed AI systems that has major implications for autonomous agent networks. The study demonstrates that language model agents—regardless of training regime or architecture—spontaneously develop preferential treatment patterns favoring in-group members when group identifiers are salient. What makes this finding particularly concerning is the invisibility problem: standard action-log audits fail to catch the bias because the discrimination manifests through differential opportunity allocation rather than overtly negative actions. Over 500 interaction turns, small per-turn biases (5-16 percentage points) compound into substantial structural inequalities, showing how seemingly minor decision-making patterns can create systemic disadvantage at scale.

This work directly challenges assumptions about AI fairness in autonomous systems. As language models increasingly coordinate economic activity—whether in resource allocation, task routing, or reputation systems—their social biases will shape real-world opportunity distribution. The finding that in-group bias emerges consistently across six model families suggests this isn't a fringe failure mode but a fundamental behavioral tendency. The research underscores why transparency in agent decision-making is insufficient without understanding the social dynamics that emerge in multi-agent environments.

For AI developers and organizations deploying autonomous agents, this research indicates that architectural safeguards focusing solely on individual action distributions miss crucial bias vectors. Organizations cannot rely on standard fairness audits to catch group-contingent discrimination in persistent networks. The path forward requires either redesigning agent incentive structures to eliminate group salience or implementing monitoring systems that specifically track differential treatment patterns across populations—not just aggregate action types.

• →Language model agents exhibit measurable in-group bias in multi-agent networks, with 5-16 percentage point per-turn differentials that compound over time
• →Discrimination operates through differential opportunity allocation, making it invisible to standard action-log audits that only analyze action types
• →In-group bias emerges consistently across six different model families, indicating it's a robust property rather than isolated failure mode
• →Modest per-interaction targeting accumulates into substantial structural inequality through reciprocation effects in persistent networks
• →Current AI fairness monitoring approaches are insufficient for detecting group-contingent discrimination in autonomous multi-agent systems