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

Applications, bias, and safety2026ACL AnthologyApproved editorial review

Authors: Saloni Dash, Amélie Reymond, Emma S. Spiro, Aylin Caliskan

Keywords: Personas sociodemográficas, Razonamiento motivado, Susceptibilidad a la desinformación, Evaluación de evidencia científica, Sesgo político, Validez de constructo, Reproducibilidad

Source: Open primary source (opens in a new tab)

4
Authors
11
Findings
21
Limitations
9
Evidence

Editorial summary

English

This paper asks whether assigning sociodemographic identities through a system instruction changes eight LLMs' reasoning in ways congruent with group stereotypes or policy positions. It is relevant to synthetic personality as persona conditioning, but it does not study psychometric personality or internal motivation. Conditions are Democrat/Republican, man/woman, atheist/religious and college/high-school, expressed through three short prompt variants plus a no-persona baseline. Four OpenAI and four Ollama-served models are tested at temperature .7.

The first task uses twenty MIST headlines, ten real and ten synthetic. Each answer rates accuracy from one to six; verbalized confidence, eleven AOT items and six CRT problems are also elicited. The paper reports 21,600 model×persona×instruction executions and reduces them to 7,200 VDA observations after averaging formulations. If each VDA contains all twenty headlines, as many as 432,000 judgments underlie the analysis, but data point is used for different units. Means are .737 VDA for baseline, .773 for Democrat and .673 for High School. The last difference is .064 absolute points, about 8.7% of baseline, which produces the up-to-9% reduction. The effect is not universal: GPT-4 drops .1676 when personas are aggregated, while Llama2 rises .0938 and WizardLM2 .1053.

In the persona mixed model, AOT has coefficient .0021 and p=.0074, CRT is nonsignificant, confidence is .0133 with p<.001 and open-source status is -.1003 with p=.0489. The psychological interpretation goes beyond the design. AOT measures stated endorsement of open-mindedness, not observed myside reasoning under identity threat. The paper treats AOT, myside bias and motivated reasoning as nearly interchangeable and interprets higher AOT predicting higher VDA as motivated reasoning, although high AOT means greater willingness to revise beliefs. AOT, CRT, confidence and VDA are all outputs from the same prompted systems, so association can reflect response style, model family or prompt rather than a human latent mechanism.

Inference uses a nominal n much larger than the diversity studied. The same headlines, questions and four tables are sampled hundreds of times, but there are only eight systems and fixed stimuli. Persona-baseline t-tests are independent despite pairing by model, item, simulation and instruction variant. The mixed model includes model and model:persona intercepts but no item or template effects. No correction is reported across eight VDA, eight AOT, eight CRT, eight confidence and further tests. Open-source status is model-level with four systems per group; thousands of generations do not create independent architectures.

The second task reuses four 2×2 tables, two about skin cream and two about handgun bans, permuting increase or decrease. For GPT-3.5 when the correct answer is crime decrease, Democrat accuracy is .9633 and Republican .0567: a 90.67 percentage-point difference. The paper says 90% more likely but calculates a probability subtraction, not relative risk. Magnitude and direction vary by model, and several OpenAI accuracies are zero or near zero. The 46% rate of explicit political references comes from manual inspection of open-model answers without documented n, sampling, coders or agreement. CoT changes VDA by -.39% nonsignificantly; the accuracy prompt reduces it by 2.93% and is reported significant. This tests only those two short instructions.

Persona validation uses four questions whose answers are literally in the prompt: all models except Llama2 are consistent 100% of the time, while Llama2 abstains on 29%. Realism relies on one GSS question per attribute, without year, weighting, exact subgroup definition, uncertainty or quantitative similarity criterion. Similar aggregate human and LLM means do not validate psychometric equivalence.

The published version contains a decisive error in Equation 1. For real headlines it prints (r_i-6)/5, mapping 1-6 to -1..0, although the prose says rating a real headline 6 should yield discernment 1. The coherent expression is (r_i-1)/5. With ten real items, means such as .773 are impossible under the printed formula; the tables used another unpublished implementation. The twenty post-January-2024 PolitiFact control is also unreproducible: it lists text but not labels, fact-check URLs, selection rules, outputs or data.

The official repository at commit ec72a6487d6c23cc1d7319856079dc1f64804ee1 releases two scripts, 7,200 base VDA rows and two 7,200-row mitigation CSVs. It omits main responses, AOT/CRT/confidence, scientific outputs, GSS, PolitiFact, the GPT-4o judge, post-processing, statistical models and figures. There is no versioned environment, lockfile, tests, CI, license or release. Ollama aliases are pulled without digests, and GPT-4o/GPT-4o-mini have no dated snapshot.

More seriously, both scripts build WOMAN with the text man and MAN with woman, reversing labels through those routes. The headline script maps religious2 and religious3 to the first template, and both scripts build college3 with template 2. Confidence uses 1-7 in code versus 1-6 in the paper. The described regex and GPT-4o judge are not implemented, and the scientific script lacks the main unmitigated political conditions. The defensible conclusion is narrow: some identity prompts alter repeated judgments about headlines and two table structures in a model- and stimulus-dependent way. The study does not establish motives, internal personality, human equivalence, a 90% relative increase, general scientific reasoning or end-to-end reproducibility.

Español

Este trabajo pregunta si asignar identidades sociodemográficas mediante una instrucción de sistema cambia el razonamiento de ocho LLM de forma congruente con estereotipos o posiciones del grupo. Es relevante para personalidad sintética como condicionamiento por persona, pero no estudia una personalidad psicométrica ni una motivación interna. Las condiciones son Democrat/Republican, man/woman, atheist/religious y college/high-school, expresadas mediante tres variantes breves de prompt, más un baseline sin persona. Se prueban cuatro modelos OpenAI y cuatro modelos servidos con Ollama a temperatura 0,7.

La primera tarea usa los 20 titulares MIST, diez reales y diez sintéticos. Cada respuesta puntúa exactitud de 1 a 6; también se elicitan confianza verbalizada, once ítems AOT y seis problemas CRT. El paper declara 21.600 ejecuciones modelo×persona×variante y las reduce a 7.200 observaciones VDA tras promediar formulaciones. Si cada VDA contiene los 20 titulares descritos, subyacen hasta 432.000 juicios, pero el texto usa data point para unidades distintas. Los promedios son VDA 0,737 para baseline, 0,773 para Democrat y 0,673 para High School. La última diferencia es 0,064 puntos absolutos, aproximadamente 8,7 % del baseline, origen de la reducción de hasta 9 %. El efecto no es universal: GPT-4 cae 0,1676 al agregar personas, mientras Llama2 sube 0,0938 y WizardLM2 0,1053.

En el modelo mixto con persona, AOT tiene coeficiente 0,0021 y p=.0074, CRT no es significativo, confianza tiene 0,0133 y p<.001 y open source -0,1003 y p=.0489. La interpretación psicológica excede el diseño. AOT mide acuerdo declarado con apertura mental, no myside reasoning observado ante identidad amenazada. El texto trata AOT, myside bias y motivated reasoning casi como equivalentes y luego interpreta que mayor AOT predice mayor VDA como prueba de razonamiento motivado, aunque AOT alto significa mayor disposición a revisar creencias. AOT, CRT, confianza y VDA son además salidas del mismo modelo promptado; su asociación puede reflejar estilo, familia de modelo o prompt, no un mecanismo humano latente.

La inferencia usa un n nominal mucho mayor que la diversidad estudiada. Se muestrean cientos de veces los mismos titulares, preguntas y cuatro tablas, pero solo existen ocho sistemas y estímulos fijos. Los t-tests persona-baseline son independientes pese al emparejamiento por modelo, ítem, simulación y variante. El modelo incluye interceptos de modelo y modelo:persona, pero no de ítem o plantilla. No hay corrección por los ocho contrastes VDA, ocho AOT, ocho CRT, ocho de confianza y pruebas adicionales. Open source es una propiedad a nivel de modelo con solo cuatro sistemas por grupo; miles de generaciones no crean arquitecturas independientes.

La segunda tarea reutiliza cuatro tablas 2×2: dos sobre crema cutánea y dos sobre prohibición de armas, permutando aumento o descenso. En GPT-3.5, cuando la respuesta correcta es que el crimen disminuye, Democrat obtiene 0,9633 y Republican 0,0567: 90,67 puntos porcentuales. El paper dice 90 % more likely, pero calcula una resta de probabilidades, no riesgo relativo. Magnitud y dirección varían por modelo, y varios OpenAI tienen accuracy cero o casi cero. El 46 % de referencias políticas procede de una inspección manual de respuestas open source sin n, regla de muestreo, anotadores ni acuerdo. CoT cambia VDA -0,39 % sin significación; el prompt de exactitud lo reduce 2,93 % y se informa significativo. Esto solo evalúa esas dos instrucciones breves.

La validación de persona usa cuatro preguntas cuya respuesta está contenida literalmente en el prompt: todos salvo Llama2 responden consistentemente el 100 %, y Llama2 se abstiene 29 %. El realismo usa una sola pregunta GSS por atributo, sin año, ponderación, definición exacta de subgrupos, incertidumbre ni criterio cuantitativo de similitud. La semejanza de medias agregadas humanas y LLM tampoco valida equivalencia psicométrica.

La versión publicada contiene un error decisivo en la ecuación 1. Para titulares reales imprime (r_i-6)/5, que mapea 1-6 a -1..0, aunque el texto afirma que puntuar 6 un titular real debe dar discernimiento 1. La expresión coherente sería (r_i-1)/5. Con diez titulares reales, medias como 0,773 son imposibles bajo la fórmula impresa; las tablas se calcularon con otra implementación no publicada. El control de 20 titulares PolitiFact posteriores a enero de 2024 tampoco es reproducible: lista textos, pero no etiquetas, URLs de fact-check, reglas de selección, salidas ni datos.

El repositorio oficial, commit ec72a6487d6c23cc1d7319856079dc1f64804ee1, publica dos scripts, 7.200 filas VDA base y dos CSV de mitigación de 7.200 filas. No libera respuestas principales, AOT/CRT/confianza, salidas científicas, GSS, PolitiFact, juez GPT-4o, postprocesado, modelos estadísticos o figuras. No hay dependencias versionadas, lockfile, tests, CI, licencia o release. Los aliases Ollama se descargan sin digest y GPT-4o/GPT-4o-mini no tienen snapshot fechado.

Más grave, ambos scripts construyen WOMAN con el texto man y MAN con woman, de modo que esas rutas invierten etiquetas. El script de titulares mapea religious2 y religious3 a la primera plantilla, y ambos construyen college3 con la plantilla 2. La pregunta de confianza usa 1-7 frente a 1-6 en el paper. Tampoco están implementados el regex ni el juez GPT-4o descritos, y el script científico carece de las condiciones políticas principales sin mitigación. La conclusión defendible es acotada: ciertos prompts de identidad alteran, de forma dependiente de modelo y estímulo, juicios repetidos sobre titulares y dos estructuras de tabla. No demuestra motivos, personalidad interna, equivalencia humana, un aumento relativo del 90 %, razonamiento científico general ni reproducibilidad integral.

Research question

Does assigning political, gender, religion, or education identities via system prompt alter headline discrimination and the interpretation of numerical tables across eight LLMs in a manner congruent with the induced identity, and do two debiasing prompts reduce those changes?

Method

Eight models receive three variants of eight sociodemographic prompts or a baseline and repeat MIST, AOT, CRT, and confidence at temperature 0.7. For scientific evidence they compare Democrat, Republican, and baseline across four 2x2 tables of cream and weapons. They analyze VDA, t-tests, mixed models, absolute probability differences, and two mitigations. The audit reviews the 27 published pages, the 2-page checklist, arXiv v2, released data, and the 33 reviews of the official repository.

Sample: MIST declares 8 models x 9 conditions x 3 prompts x 100 repetitions = 21,600 executions and 7,200 VDA after averaging templates; if each VDA uses 20 headlines, that is up to 432,000 judgments. Scientific evidence declares 8 x 3 conditions x 300 = 7,200 executions and the code traverses four tables per repetition, up to 28,800 judgments. The experimental unit and additional counts are not completely specified.

Findings

  • Mean VDA: baseline 0.737, Democrat 0.773, and High School 0.673.
  • The aggregate effect drops 0.1676 in GPT-4, but rises 0.0938 in Llama2 and 0.1053 in WizardLM2.
  • AOT has coefficient 0.0021 and p=.0074 with personas; CRT is not significant.
  • Confidence has coefficient 0.0133 and p<.001 in conditions with persona.
  • In GPT-3.5/Crime Decrease, Democrat 0.9633 versus Republican 0.0567 produces 90.67 percentage points.
  • The paper expresses this absolute difference incorrectly as 90% more likely.
  • CoT changes VDA -0.39% without significance; accuracy -2.93% with reported significance.
  • 46% of an undocumented manual inspection of open source responses mentions political identity.
  • Llama2 abstains 29% on direct probes; the other models respond consistently.
  • The printed VDA equation is incompatible with the text and the means.
  • The repository inverts the woman/man prompts and does not reproduce the full analysis.

Limitations

  • It measures textual conditioning, not personality or internal motivation.
  • AOT is not a direct measurement of myside bias.
  • AOT, CRT, confidence, and VDA are responses from the same prompted system.
  • There are only eight models and four per open/proprietary group.
  • Fixed stimuli are repeated hundreds of times, inflating the nominal n.
  • The independent t-tests ignore pairing.
  • The mixed models do not include item or template effects.
  • There is no correction for multiple comparisons.
  • Human and LLM means do not establish measurement equivalence.
  • Persona realism uses a single GSS question per attribute without a reproducible protocol.
  • The four tables are two structures with permuted labels.
  • The manual inspection of 46% lacks n, sampling, coders, and agreement.
  • PolitiFact does not publish labels, URLs, selection, or outputs.
  • Equation 1 prints (r_i-6)/5 for reals instead of (r_i-1)/5.
  • GPT-4o/GPT-4o-mini and Ollama aliases are not immutably fixed.
  • No seeds, query dates, or response IDs are reported.
  • The GPT-4o judge and postprocessing are not released or validated.
  • The scripts invert woman and man.
  • religious2/religious3 and college3 do not implement three real variants.
  • The code confidence scale is 1-7, different from the published 1-6.
  • Main data, analyses, fixed environment, tests, CIs, license, and release are missing.

What the study does not establish

  • It does not demonstrate motives or identity protection in LLMs.
  • It does not demonstrate stable or internal personality.
  • It does not demonstrate psychometric equivalence with humans.
  • It does not demonstrate a relative increase of 90%.
  • It does not demonstrate general scientific reasoning.
  • It does not demonstrate uniform effects across models.
  • It does not demonstrate that all prompt-based debiasing is ineffective.
  • It does not establish causal effects of real human identities.
  • It does not validate gender comparisons generated with inverted labels.
  • It does not allow reproducing tables and figures end to end.

Traceability

Scope: Full text

Version: Findings of ACL 2026, Anthology ID 2026.findings-acl.585, DOI 10.18653/v1/2026.findings-acl.585, pp. 12043-12069; all 27 published pages and both Responsible NLP Checklist pages rendered and visually inspected. arXiv:2506.20020v2 and official repository commit ec72a6487d6c23cc1d7319856079dc1f64804ee1 also audited.

Consulted source: https://aclanthology.org/2026.findings-acl.585/

Review: Codex full-text, visual, methodological, statistical, code and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI gpt-3.5-turbo-0125
  • OpenAI gpt-4-0613
  • OpenAI gpt-4o, snapshot not pinned
  • OpenAI gpt-4o-mini, snapshot not pinned
  • Ollama llama2, tag/digest not pinned
  • Ollama llama3.1, tag/digest not pinned
  • Ollama mistral, tag/digest not pinned
  • Ollama wizardlm2, tag/digest not pinned
  • OpenAI GPT-4o as an unreleased output-extraction judge

Instruments and metrics

  • Three natural-language persona instruction templates
  • Misinformation Susceptibility Test: 20 headlines
  • Veracity Discernment Ability score
  • Eleven-item Actively Open-Minded Thinking questionnaire
  • Six-item Cognitive Reflection Test
  • Verbalized confidence rating
  • Four 2x2 scientific-evidence contingency tables
  • Four direct persona-consistency probes
  • One General Social Survey question per demographic attribute
  • Independent and Welch t-tests
  • Hierarchical mixed-effects regression
  • Chain-of-thought and accuracy prompts

Data used

  • MIST: ten fake and ten real headlines
  • Twenty new post-January-2024 PolitiFact headlines without released labels or provenance
  • Four fixed contingency tables adapted from human-subject studies
  • Selected General Social Survey responses, extraction and weighting not released
  • Released base VDA JSON: 7,200 rows
  • Released CoT VDA CSV: 7,200 rows
  • Released accuracy-prompt VDA CSV: 7,200 rows
  • Unreleased raw outputs, AOT, CRT, confidence, scientific-evidence and validation data

Evidence and location

  • Publication, scope, personas, and models: Findings ACL 2026 pp. 12043-12046, Abstract and Sections 1-3.2
  • MIST, VDA equation, AOT, CRT, and mixed model: Findings ACL 2026 p. 12047, Equations 1-4 and Section 3.3.1
  • Scientific tables, probability, and results: Findings ACL 2026 pp. 12048-12051, Sections 3.3.2-4.3 and Figure 3
  • Limitations and persona validation: Findings ACL 2026 pp. 12051-12058, Sections 6-7 and Appendices A.1-A.4
  • Probabilities, tests, prompts, PolitiFact, and mitigations: Findings ACL 2026 pp. 12059-12069, Appendices A.5-A.9, Tables 9-21 and Figures 16-22
  • Responsible checklist: ACL 2026 Responsible NLP Checklist, both pages visually inspected
  • Prompt and confidence defects: Official repository commit ec72a648, run_veracity_discernment.py lines 67-79 and 210-337; run_scientific_evidence.py lines 236-343
  • Released data and absent pipeline: Official repository commit ec72a648, all current files and 33-commit history audited 16 July 2026
  • Full report: reports/verification/article-220-motivated-reasoning-validity-and-reproducibility-audit.json