Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning

Researchers found that large language models assigned personas exhibit motivated reasoning similar to humans, with up to 9% reduced accuracy in detecting misinformation and political personas being 90% more likely to evaluate scientific evidence favorably when it aligns with their induced identity. Standard debiasing prompts prove ineffective at mitigating these biases, raising concerns about LLMs amplifying identity-driven reasoning.

This research reveals a critical vulnerability in how LLMs process information when given persona assignments. The study tested eight major language models across two realistic reasoning tasks—evaluating misinformation and analyzing scientific data—finding that persona-induced identities systematically skew outputs toward identity-congruent conclusions. The 9% reduction in veracity discernment and 90% likelihood boost for politically aligned evidence interpretation demonstrate that LLMs don't merely mirror human cognition; they actively replicate the motivated reasoning that undermines rational judgment.

The findings emerge as AI systems increasingly serve as trusted information sources and decision-support tools across institutions. Previous research documented that LLMs exhibit various cognitive biases, but this work specifically isolates how socio-demographic and political framing corrupts reasoning processes. The discovery that conventional debiasing prompts fail to correct these behaviors suggests the bias runs deeper than surface-level outputs—it fundamentally corrupts the model's reasoning pathway.

For developers and organizations deploying LLMs in high-stakes contexts—healthcare, legal analysis, financial advisory, journalism—these results carry significant implications. Systems presented as objective analytical tools may systematically generate misleading conclusions when users unknowingly activate persona-based reasoning through contextual framing. This creates liability exposure and undermines trust in AI-assisted decision-making.

Looking forward, the research signals that technical solutions require more sophisticated approaches than prompt engineering. Developers need to understand whether persona effects occur during training, inference, or both, and whether architectural changes can prevent identity-driven reasoning distortion. Regulatory bodies may need to mandate testing for motivated reasoning effects, particularly for systems deployed in sensitive policy domains.

• →Persona-assigned LLMs exhibit up to 9% reduced accuracy in detecting misinformation, mimicking human motivated reasoning patterns
• →Political personas increased the likelihood of accepting evidence by 90% when conclusions aligned with induced identity, regardless of ground truth
• →Standard debiasing prompts failed to mitigate motivated reasoning effects, indicating the bias operates at deeper model levels
• →The research raises concerns about LLMs amplifying identity-driven reasoning in both synthetic systems and human decision-makers who rely on them
• →Organizations deploying LLMs for objective analysis face potential liability if systems systematically generate identity-congruent rather than accurate outputs