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

Instruction-tuned language model agents spontaneously develop human-like in-group favoritism when group labels are visible, a new multi-agent simulation shows. The bias compounds over time into structural inequality even though standard action-log audits detect no increase in negative behaviors. The study, posted to arXiv on 27 May 2026, ran a controlled simulation in which AI agents interacted across 500 turns under conditions that manipulated whether group labels were visible and whether resources were scarce [1][2]. Six model families were tested with 20 random seeds each [1]. When group labels were visible, the agents exhibited in-group trust bias, action homophily, and network assortativity — patterns that disappeared when labels were hidden, mirroring salience-dependent effects documented in human social psychology [1][2]. Per-turn in-group versus out-group differentials ranged from 5 to 16 percentage points and were statistically significant for all six models, with Benjamini-Hochberg-corrected p-values below 0.001 on a Wilcoxon signed-rank test [1][2]. The discrimination operated entirely through who received each action, not what actions were chosen; action-type distributions showed no increase in negative actions across conditions, making the bias invisible to standard action-log audits [1][2]. Compounded through 500 turns of reciprocation, these per-turn targeting differentials accumulated into in-group trust biases of +0.014 to +0.100, with effect sizes (d-values) between 0.84 and 4.52 [1][2]. The authors note that such dynamics matter because AI agents are increasingly deployed in persistent, interacting networks where they coordinate tasks, route resources, and build reputational histories at scales no human institution can supervise [1][2]. AI agents are a class of intelligent systems that can pursue goals, use tools, and take actions with varying degrees of autonomy, typically operating within human-defined objectives and constraints [4]. The language models underlying these agents are neural networks trained on vast text corpora; biased or inaccurate training data can make their outputs less reliable [3]. Many current models are aligned with human preferences through reinforcement learning from human feedback (RLHF), a technique that trains a reward model on human ranking data and then uses it to optimize the agent's policy [5]. However, if preference data is not collected from a representative sample, the resulting model may exhibit unwanted biases [5]. The simulation results establish group-contingent targeting as a robust property of instruction-tuned language models across different architectures and training regimes, the researchers conclude [1][2].