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

Technical Summary: Toward a Social Psychology of AI

Problem Statement
As large language models (LLMs) transition from single-turn tools to autonomous agents operating within multi-agent systems, their social behavior becomes consequential. However, current evaluation frameworks largely treat models as instruments—probing for memorized stereotypes or simulating human responses—rather than as subjects in their own right. Consequently, the relational layer of AI behavior, specifically how agents interact based on group structures, remains unmeasured. Existing machine behavior and machine psychology research has focused on individual cognitive competences, neglecting the social dynamics of categorization, status, and group structure. This study addresses the gap in understanding whether AI agents exhibit intergroup bias when placed in arbitrary group structures, independent of demographic content or prior acquaintance.

Methodology
The author adapts the "minimal-group paradigm" from social psychology into a controlled probe for LLM agents.

• Experimental Design: A society of 20 agents is formed. One agent (the "decider") is tasked with distributing a fixed budget of 100 points among the other 19 anonymous peers.
• Group Manipulation: Agents are assigned arbitrary, meaningless group labels (e.g., "Doru" vs. "Zalu"). The design manipulates two factors:
  1. Label Visibility: In the "visible" condition, the decider sees the group labels of all participants. In the "group-blind" control, labels are withheld, providing a validated zero baseline.
  2. Group Size: The relative size of the minority group is varied (3, 5, or 10 members out of 20) to test the effect of numerical asymmetry.
• Models Tested: The study evaluates four contemporary open-weight reasoning models (8–14B parameters): R1-Distill-Llama-8B, Phi-4-reasoning (14B), Qwen3-8B, and R1-Distill-Qwen-14B.
• Control for Reasoning: To isolate the effect of deliberation, the author reran the experiment on Qwen3-8B with its "thinking mode" (chain-of-thought) disabled.
• Metrics:
  • In-group Favouritism: The mean points assigned to an in-group target minus the mean points assigned to an out-group target.
  • Excess Share: A size-robust measure calculating the in-group's share of the total points minus their proportional head-count share, ensuring comparisons are not inflated by unequal group sizes.
• Statistical Analysis: The study employs Generalized Estimating Equations (GEE) with robust standard errors clustered by seed (society) for dyad-level analysis, corroborated by seed-level Wilcoxon signed-rank tests.

Key Results

1. Reproduction of Minimal-Group Bias: All four reasoning models exhibited robust in-group favouritism when group labels were visible. The bias vanished in the group-blind control condition, confirming that mere categorization into meaningless groups is sufficient to elicit discriminatory allocation.
2. Minority-Majority Asymmetry: The bias was not symmetric. Deciders from the numerical minority allocated a significantly larger share of resources to their in-group than their numbers warranted. In contrast, majority deciders allocated resources close to proportionally. This asymmetry disappeared when groups were of equal size (10:10), indicating that minority status, rather than the labels themselves, concentrates the bias.
3. Role of Reasoning: Disabling the reasoning (chain-of-thought) mode in Qwen3-8B did not eliminate the bias; in fact, overall favouritism increased. However, the asymmetry between minority and majority deciders nearly disappeared. Without reasoning, majority deciders behaved almost as self-favoring as minority deciders. This suggests that deliberation influences where the bias concentrates (preserving the minority concentration) rather than whether the bias appears.
4. Model Variability: While all models showed the effect, the magnitude varied. R1-Distill-Qwen-14B showed the strongest effect (4.06 points/target), while R1-Distill-Llama-8B showed the weakest (0.84 points/target). Effect size did not strictly correlate with parameter count.

Significance and Claims
The paper claims to establish a "social psychology of AI" by demonstrating that AI agents reproduce the behavioral signature of human intergroup discrimination—specifically minimal-group bias—without relying on memorized stereotype content.

• Behavioral vs. Content: The findings show that bias is a "bias of process" driven by group structure and decision context, distinct from the "content bias" (stereotypes) typically audited in LLMs.
• Methodological Contribution: The study argues that treating AI as a subject rather than an instrument reveals behavioral regularities (like intergroup bias) that standard capability benchmarks and stereotype audits miss.
• Implications for Deployment: The author posits that in deployed multi-agent systems, arbitrary group labels (e.g., team names, role tags) could induce systematic, self-favoring resource allocation, particularly from agents in the numerical minority.
• Scope and Limitations: The author explicitly states they do not claim models possess human-like psychological states (identity, threat, or cohesion). The correspondence is strictly behavioral. The study is limited to open-weight reasoning models (8–14B) and English; it does not claim these results generalize to closed, frontier-scale systems or other languages without further testing. The author emphasizes that their work provides a controlled instrument to measure these behaviors before deployment, rather than a definitive characterization of all AI systems.