Give LLM agents something to lose and group favoritism shifts from the label to history: 3.3 million calls show conformity to observed norms
Across fifteen OpenAI and three Claude models, agents in small arbitrarily labeled societies played ten rounds of point sharing and then faced one-shot decisions differing in exactly one respect. A bare label produced 6–9 points of favoritism only when giving was costless; once agents could keep points it fell to about 2, history became the main source (positive on 13 of 15 models, 3.5–8 points on 11, transferring to labeled strangers), and scripted histories showed egalitarian norms zeroing favoritism, counter-norms reversing it, and stronger models siding with an individual's record.