AI agents replicated human group favoritism

A preprint paper by Messi H. J. Lee (arXiv:2609.00009) published on arXiv shows that AI agents from four open-weight reasoning models replicate human minimal-group favoritism: an arbitrary 'in-group' label causes bias in the distribution of 100 points, which disappears in a group-blind control and concentrates among the numerical minority.

Compiled by Sergey KostenchukPublished 2026-09-07Updated 2026-09-18

2026-09-07 Research Meta

🤖 AI agents replicated human group favoritism

A paper titled 'Toward a social psychology of AI' (arXiv:2609.00009) was published on arXiv: an agent in a society of 20 agents divides 100 points among 19 anonymous participants based on an arbitrary group label. In four open-weight reasoning models (Qwen3-8B, DeepSeek-R1-Distill, Phi-4-reasoning), the 'in-group' label caused favoritism — it disappeared in a group-blind control and was strongest among decision-makers from the minority.

🌍 Standard LLM audits look for learned stereotypes but do not measure agent behavior in groups. The test gives red-teamers of multi-agent systems a portable tool: bias without stereotypes in the data.

👤 If you run agents in groups — keep in mind: they favor a meaningless 'in-group' label, and turning off reasoning changes where the bias concentrates. The paper and models are open — the test can be replicated.

Source 1: https://arxiv.org/abs/2609.00009 Source 2: https://arxiv.org/html/2609.00009v1

Sources