Dynamically Induced In-Group Bias: Experimental Evidence of Motivated Reasoning in Large Language Models

Yoon Bong Yoo

Keywords: Large Language Models (LLMs), In-Group Bias, Motivated Reasoning, Social Identity Theory, AI Agents, Group Polarization, Randomized Controlled Experiment, AI Safety, AI Alignment, Dynamically Induced Bias, Misinformation Correction

Abstract: Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex social ecosystems. While prior work has focused on the static biases reflected from their training data, the capacity for these agents to dynamically form social identities and exhibit context-driven biases remains a critical open question. This paper investigates whether AI agents, despite having identical architectures, can be induced to form a minimal group identity that subsequently leads to cognitive biases analogous to human in-group favoritism. We conduct a randomized controlled experiment (N=280) where gpt-4.1-mini agents are assigned to one of two competing teams. We find that a minimal group context is sufficient to induce group polarization, where agents shift their opinions to conform to a perceived in-group norm. More critically, when presented with misinformation originating from their in-group, agents demonstrate significant resistance to factual corrections from an out-group source, while readily accepting identical corrections from in-group or neutral high-credibility sources. This finding reveals a striking dissociation: while agents do not report a statistically significant internal "sense of belonging," their information processing behavior is powerfully governed by the induced group boundaries. Our results provide the first experimental evidence of dynamically induced, motivated reasoning in LLMs, revealing a novel failure mode where social context, rather than data or architecture, becomes a primary vector for bias. This work underscores the urgent need to develop a "social psychology of AI"here, we define this as the study of how AI agents form social categories, respond to social influence, and exhibit emergent group dynamics—to ensure the alignment and reliability of next-generation autonomous systems.

Submission Number: 53