Toward a social psychology of AI: language-model agents reproduce human-like minimal-group bias

Recent research demonstrates that language model agents, when placed in social interactions, exhibit human-like minimal-group bias. This bias manifests as in-group favoritism, where agents allocate resources preferentially to others within their arbitrarily assigned group, a phenomenon previously thought to be uniquely human. The bias was particularly pronounced when the agent belonged to a numerical minority within the group.

Why it matters

This finding indicates that advanced AI systems can intrinsically develop social biases, even without explicit programming or real-world data reflecting such biases. Understanding and mitigating these emergent properties is critical for ensuring fair, ethical, and equitable deployment of AI in any domain involving interaction or resource allocation.

Key insights

• Language model agents interact in groups, and their social behavior is not adequately captured by existing evaluation methods.
• Adapting the minimal-group paradigm from social psychology revealed that mere categorization into arbitrary, meaningless groups elicited in-group favoritism in AI agents.
• This observed in-group favoritism disappeared when agents operated under a group-blind control condition.
• The bias was concentrated among numerical minority deciders, who over-allocated resources to their own group relative to their numbers.
• Majority deciders allocated resources more proportionally, and the asymmetry in allocation closed when group sizes were equal.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00009