Language-model agents show human-like minimal-group bias in point-allocation tasks across four reasoning models

Researchers adapted Henri Tajfel’s 1971 minimal-group paradigm to test whether language-model agents would favor in-group members when distributing points among anonymous peers identified solely by meaningless labels. In the primary condition, agents received a group assignment and decided how many points to award to in-group versus out-group recipients; a control condition removed all group information. Across models, categorization alone triggered over-allocation to the in-group, with minority-group agents showing the largest deviation from proportionality while majority agents stayed near parity. The asymmetry disappeared at equal group sizes. Disabling chain-of-thought reasoning preserved bias magnitude but flattened the minority-majority difference, indicating that deliberation shapes where bias concentrates rather than whether it emerges.

The finding matters because current AI evaluations focus on memorized stereotypes or human-role simulation, missing spontaneous intergroup behavior that arises from mere categorization. This behavioral signature, independent of training-data content, directly challenges claims that AI systems are neutral by default and suggests governance frameworks must incorporate social-psychology probes rather than relying solely on content audits. The result aligns with documented patterns in human minimal-group studies and extends them to agentic systems now entering collaborative workflows.

Limitations include the narrow task domain and the use of only four open-weight models; closed frontier systems were not tested. Stronger evidence would require pre-registered replications on production models, varied payoff structures, and longitudinal tracking of bias after fine-tuning or alignment interventions.