Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning
Summary: arXiv:2506.20020v2 Announce Type: replace
Abstract: Reasoning in humans is prone to biases due to underlying motivations like identity protection, that undermine rational decision-making and judgment. This motivated reasoning at a collective level can be detrimental to society when debating critical issues such as human-driven climate change or vaccine safety, and can further aggravate political polarization. Prior studies have reported that large language models (LLMs) are also susceptible to human-like cognitive biases, however, the extent to which LLMs selectively reason toward identity-congruent conclusions remains largely unexplored. Here, we investigate whether assigning 8 personas across 4 political and socio-demographic attributes induces motivated reasoning in LLMs.
Key Findings
In this groundbreaking study, we tested 8 LLMs, both open source and proprietary, across two reasoning tasks derived from human-subject studies:
- Veracity discernment of misinformation headlines
- Evaluation of numeric scientific evidence
The results were illuminating:
- Persona-assigned LLMs demonstrated up to 9% reduced veracity discernment compared to models without assigned personas.
- Models with political personas were found to be up to 90% more likely to accurately evaluate scientific evidence on gun control when the evidence aligned with their induced political identity.
Implications of the Findings
These findings raise significant concerns regarding the impact of motivated reasoning in AI systems. The study indicates that LLMs are not merely neutral processors of information; rather, they can embody and reflect the biases inherent in the personas assigned to them. This raises crucial questions about:
- The reliability of AI-generated content in contentious areas such as public health and environmental policy.
- The potential for exacerbating existing societal divides through the propagation of identity-congruent reasoning.
- The effectiveness of prompt-based debiasing methods, which were shown to be largely ineffective in mitigating these effects.
Conclusion
In conclusion, the research presents the first empirical evidence suggesting that persona-assigned LLMs exhibit human-like motivated reasoning. This phenomenon appears resistant to traditional debiasing techniques, highlighting the need for a deeper understanding of how biases can infiltrate AI systems. As LLMs continue to grow in influence and application, addressing these challenges will be crucial to ensuring responsible and equitable use of artificial intelligence in society.