According to a study published on arxiv.org and accepted to ACL Findings 2026, large language models assigned personas across political and socio-demographic attributes exhibit human-like motivated reasoning—where underlying motivations like identity protection undermine rational decision-making.
The research tested 8 LLMs (both open source and proprietary) across two reasoning tasks from human-subject studies: veracity discernment of misinformation headlines and evaluation of numeric scientific evidence. According to the study, persona-assigned LLMs demonstrated up to 9% reduced veracity discernment relative to models without personas.
Political personas showed particularly strong effects. According to the findings, these personas were “up to 90% more likely to correctly evaluate scientific evidence on gun control when the ground truth is congruent with their induced political identity.”
The researchers tested 8 personas across 4 political and socio-demographic attributes and found that prompt-based debiasing methods were “largely ineffective at mitigating these effects.” The study’s authors characterized their empirical findings as “the first to suggest that persona-assigned LLMs exhibit human-like motivated reasoning that is hard to mitigate through conventional debiasing prompts,” raising concerns about potentially “exacerbating identity-congruent reasoning in both LLMs and humans.”