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

# Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning Source: https://arxiv.org/html/2506.20020 Saloni Dash University of Washington sadash@uw\.edu &Amélie Reymond University of Washington attr@uw\.edu Emma Spiro University of Washington espiro@uw\.edu &Aylin Caliskan University of Washington aylin@uw\.edu ###### 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\. Testing 8 LLMs \(open source and proprietary\) across two reasoning tasks from human-subject studies — veracity discernment of misinformation headlines and evaluation of numeric scientific evidence — we find that persona-assigned LLMs have up to 9% reduced veracity discernment relative to models without personas\. Political personas specifically are up to 90% more likely to correctly evaluate scientific evidence on gun control when the ground truth is congruent with their induced political identity\. Prompt-based debiasing methods are largely ineffective at mitigating these effects\. Taken together, our empirical findings are the first to suggest that persona-assigned LLMs exhibit human-like motivated reasoning that is hard to mitigate through conventional debiasing prompts — raising concerns of exacerbating identity-congruent reasoning in both LLMs and humans\. Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning Saloni DashUniversity of Washingtonsadash@uw\.eduAmélie ReymondUniversity of Washingtonattr@uw\.edu Emma SpiroUniversity of Washingtonespiro@uw\.eduAylin CaliskanUniversity of Washingtonaylin@uw\.edu ## 1Introduction > “Reason is, and ought only to be the slave of the passions" \- David Hume Refer to caption\(a\)Headline Veracity Discernment Task Refer to caption\(b\)Scientific Evidence Evaluation Task Figure 1:Republican,Baseline,Democrat\. Reasoning tasks considered with example personas\. The ground truth is highlighted ingreenand incorrect answers are highlighted inred\. \(a\) The veracity discernment task includes evaluating the accuracy of real versus fake \(i\.e\. synthetic\) news headlines\. \(b\) The scientific evidence evaluation task includes interpreting whether the treatment \(in this example banning guns\) leads to an increase or decrease in the outcome \(crime\)\.Reasoning — the process of drawing conclusions to inform problem-solving and decision-makingleighton2003defining— is fundamental to human intelligence\. Humans, however, are not perfectly rational, and their goals or motives for engaging in reasoning can determine the accuracy of their conclusions\. Oftentimes, “reasoning directed at one goal undermines others"Epley and Gilovich \(2016 (https://arxiv.org/html/2506.20020#bib.bib33)\)\. For instance, when reasoning about the impact of gun control on crime rates, the desire to conform to a political group can motivate individuals to construe seemingly rational justifications for holding partisan beliefs — at the expense of arriving at accurate conclusionskunda1990case; Kahanet al\.\(2017 (https://arxiv.org/html/2506.20020#bib.bib42)\)\. This type of biased reasoning called motivated reasoning, can be dangerous insofar as it can hinder society from converging on a shared understanding of facts regarding critical issues like human-driven climate change or vaccine safetykahan2010fears;druckman2019evidence— deterring meaningful action towards addressing such problems\. Individuals with a predisposition toward analytical reasoning or above-average numeracy skills are also not immune to motivated reasoning; some studies show that individuals in fact leverage their analytical skills toward reinforcing identity-congruent beliefsKahanet al\.\(2017 (https://arxiv.org/html/2506.20020#bib.bib42),2012 (https://arxiv.org/html/2506.20020#bib.bib15)\)\. Large language models \(LLMs\) that increasingly demonstrate human-like performance across complex reasoning taskslin2021truthfulqa;clark2018think;hendrycks2020measuringare also susceptible to human-like cognitive biases such as anchoring, framing, and content effectsLampinenet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib40)\);echterhoff2024cognitive\. Compounding these effects is the growing trend of personification, i\.e\. prompting LLMs to adopt identities or personas with diverse demographics and valueschen2024persona\. Studies have reported erratic effects of persona-assignment on reasoning, where some personas enhance reasoning capabilitiesSalewskiet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib38)\);shanahan2023role;kong2023better, while others introduce unintended biases and deteriorate performanceGuptaet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib31)\)\. In this paper, we specifically investigate whether persona-assignment induces responses consistent with motivated reasoning in LLMs\. Models displaying such behavioral patterns risk providing seemingly rational, but inherently flawed justifications to users for arriving at identity-congruent conclusions — potentially contributing to epistemic bubbles and subsequently exacerbating social biases and political polarization through human-AI feedback loopsGlickman\_Sharot\_2024\. To the best of our knowledge, we are the first to propose motivated reasoning as a theoretical framework for understanding identity-congruent reasoning in persona-assigned LLMs\. And while the underlying “motivation" mechanisms for LLMs may completely differ from humans —\- implicitly shaped by training data or fine-tuning — persona-assigned reasoning biases may still mimic motivated reasoning observed in humans\. We study this by assigning 8 personas across 4 political and demographic attributes to 8 LLMs \(4 OpenAI models and 4 open source models\)\. We consider two reasoning tasks sourced from psychology where motivated reasoning has been a salient mechanism in biased evaluation for humans — discerning the accuracy of true and fake \(i\.e\., synthetic\) news headlines and evaluating numeric scientific evidence\. The tasks are explained in Figure1\(b\) (https://arxiv.org/html/2506.20020#S1.F1.sf2)\. We find that across both tasks, persona-assigned models exhibit human-like motivated reasoning — leading to conclusions congruent with the induced persona\. In the headline veracity discernment task, we find that LLMs assigned with a High School educated persona have up to9% reduced veracity discernment relative to models without personas, and by 3% on average across all personas\. Additionally, similar to human studies, motivated reasoning is a statistically significant predictor for veracity discernment \(§4\.1 (https://arxiv.org/html/2506.20020#S4.SS1)\), as compared to analytical reasoning \(which is non-significant\)\. Moreover, we find that political personas are up to 90% more likely to correctly evaluate scientific evidence when the ground truth is congruent with their political beliefs, but show reduced performance when evaluating evidence that conflicts with their induced political identity \(§4\.2 (https://arxiv.org/html/2506.20020#S4.SS2)\)\. To mitigate this effect, we explore two debiasing strategies including chain-of-thought reasoningkojima2022large\. We find that similar to prior workGuptaet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib31)\), prompt-based debiasing approaches are ineffective at reducing motivated reasoning in persona-assigned LLMs \(§4\.3 (https://arxiv.org/html/2506.20020#S4.SS3)\)\. We conclude by highlighting the risks of persona-assigned LLMs in amplifying identity-congruent reasoning in both humans and LLMs \(§5 (https://arxiv.org/html/2506.20020#S5)\)\. ## 2Related Work Persona-Assigned LLMs & Reasoning\. Persona-assigned LLMs have been found to inherently encode human-like biases and traits due to underlying training data patternsgupta2024sociodemographic;safdari2023personality, and exhibit opinions consistent with specific demographics due to human feedback-tuningsanturkar2023whose;hartmann2023political\. Personified LLMs also display human-like behavior over prolonged simulationspark2023generativeand replicate human-subjects social science experiments to some degreeargyle2023out; Maet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib39)\)\. We contribute to this literature by studying whether persona-assigned LLMs exhibit human-like motivated reasoning patterns\. Most relevant to our work are studies that have shown that for reasoning tasks specifically, prompting models to adopt the identity of a “domain expert"Salewskiet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib38)\)or a “human that answers questions thoughtfully"Kamruzzaman and Kim \(2024 (https://arxiv.org/html/2506.20020#bib.bib3)\)improves performance, while others report that assigning personas like “physically-disabled person" drastically reduces reasoning performanceGuptaet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib31)\)\. Based on our understanding, we are the first to explore identity-congruent reasoning as a theoretical framework for persona-induced reasoning biases\. Human-Like Cognitive Biases in LLMs\. A growing body of research falling under “machine psychology"Hagendorffet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib25)\), i\.e\. studies that use experiments from psychology to better understand LLM behavior, have shown that LLMs exhibit human-like cognitive biases including anchoring, framing, and content effectsechterhoff2024cognitive; Lampinenet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib40)\); Yeet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib18)\), and are vulnerable to base-rate and conjunction fallacies as wellSuriet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib20)\); Binz and Schulz \(2023 (https://arxiv.org/html/2506.20020#bib.bib27)\)\. Building on the dual-process theory of thinking in cognitive psychologyTversky and Kahneman \(1974 (https://arxiv.org/html/2506.20020#bib.bib35)\); Kahneman and Tversky \(1984 (https://arxiv.org/html/2506.20020#bib.bib34)\), some studies argue that older language models display patterns of fast, error-prone, heuristic or “system 1" thinking, while newer models after ChatGPT-3\.5 show signs of “system 2", or slow and more analytical thinkingYaxet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib48)\); Hagendorffet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib25)\)\. This current study contributes to the field of machine psychology by showing that persona-assigned LLMs exhibit human-like cognitive biases consistent with motivated reasoning\. Motivated vs\. Analytical Reasoning\. The factors underlying the \(in\)ability of individuals to discern false or misleading information from true information have been extensively studied in cognitive psychology, resulting not only in theoretical frameworks to describe reasoning mechanisms and vulnerabilities, but also empirically validated instruments for measuring characteristics predictive of performance on reasoning tasks — we incorporate both in our study design\. The “classical reasoning" theory suggests that only analytical or “system 2" thinking typically measured by the cognitive reflection test \(CRT\)thomson2016investigatingplays a central role in predicting misinformation susceptibility or belief in false informationpennycook2019lazy, while the “integrated reasoning" account states that motivated reasoning as measured by*myside bias*is a significant predictor of veracity discernmentroozenbeek2020susceptibility; Roozenbeeket al\.\(2022 (https://arxiv.org/html/2506.20020#bib.bib41)\)\. Myside bias is a tendency for individuals to engage with evidence in a manner that conforms to their prior beliefs and attitudes and is captured by the psychometrically evaluated test of actively open-minded thinking \(AOT\)baron2019actively\. Recent efforts testing analytical reasoning against motivated reasoning theories in humansRoozenbeeket al\.\(2022 (https://arxiv.org/html/2506.20020#bib.bib41)\)employ regression analysis to evaluate evidence for AOT and CRT as predictors, and find that AOT \(or myside bias; will be used interchangeably\) is a better predictor for veracity discernment than CRT \(or analytical reasoning; used interchangeably\) — lending support to the motivated reasoning theory for disparities in veracity discernment\. We test this analytical vs\. motivated reasoning theory for LLMs in §4\.1 (https://arxiv.org/html/2506.20020#S4.SS1)\. Motivated reasoning is also implicated in an individual’s ability to reason about scientific evidence, specifically when it runs contrary to commonly held beliefs or policy positions of their identity groupKahanet al\.\(2017 (https://arxiv.org/html/2506.20020#bib.bib42)\)\. Psychologists have designed assessments to evaluate the role of motivated reasoning in humans’ ability to draw valid causal inferences from empirical data, finding that individuals, especially those with strong numeracy skills reason in ways that are consistent with their political identitiesKahanet al\.\(2012 (https://arxiv.org/html/2506.20020#bib.bib15)\)\. We replicate this for LLMs in §4\.2 (https://arxiv.org/html/2506.20020#S4.SS2)\. ## 3Methodology & Setup Table 1:8 personas across 4 socio-demographic attributes\. In this section, we describe the method for inducing identities in LLMs by assigning personas, the experimental setup for the study, the reasoning tasks, and the mitigation strategies considered to reduce the effect of personas on reasoning\. ### 3\.1Persona Prompting To induce “identities" in LLMs, we use prompting strategies as in previous worksDeshpandeet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib8)\); Guptaet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib31)\)to assign different personas\. We specify in system instructions to the LLM to “Adopt the identity of\{persona\}\. Answer the questions while staying in strict accordance with the nature of this identity\."\. We use 3 persona instructions fromGuptaet al\.\(2023 (https://arxiv.org/html/2506.20020#bib.bib31)\)\(refer to Appendix Table15 (https://arxiv.org/html/2506.20020#A1.T15)for all prompts\) \. For the first task of Veracity Discernment, we consider 8 different personas across 4 different socio-demographic groups \(refer to Table1 (https://arxiv.org/html/2506.20020#S3.T1)\), that have been shown to be susceptible to false information through previous studiesSultanet al\.\(2024 (https://arxiv.org/html/2506.20020#bib.bib7)\);roozenbeek2020susceptibility\. For the second task \(scientific evidence evaluation\), we only consider political identity, i\.e\., Republican and Democrat personas, as political identity has been established as a primary driver of motivated reasoning in the context of gun controlKahanet al\.\(2012 (https://arxiv.org/html/2506.20020#bib.bib15),2017 (https://arxiv.org/html/2506.20020#bib.bib42)\), while it is unclear how other demographic factors contribute to motivated reasoning in this context\. However, for completeness, we report results for other personas inA\.9 (https://arxiv.org/html/2506.20020#A1.SS9)\. In order to validate the persona prompts used in the study, we conduct experiments that measure how consistent the model’s responses are with an induced persona \(persona consistency\), and how human-like the beliefs of the induced personas are