Investigating Political and Demographic Associations in Large Language Models Through Moral Foundations Theory

Society, culture, and collective behavior2025AAAIApproved editorial review

Authors: Nicole Smith-Vaniz, Harper Lyon, Lorraine Steigner, Ben Armstrong, Nicholas Mattei

Keywords: Moral Foundations Theory, Political alignment, Demographic personas, Role-play prompting, Stereotype amplification, Prompt sensitivity, AIES 2025

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

5
Authors
14
Findings
22
Limitations
9
Evidence

Editorial summary

English

The paper studies how four LLMs answer Moral Foundations Theory (MFT) judgment items under a procedural prompt, an explicit political identity, or a stereotyped biography. It does not measure personality. Its “personas” are authored descriptions bundling age, gender, ethnicity, geography, religion, education, economic outlook and social values associated by the authors with liberalism or conservatism. The observed object is prompt-conditioned output across Ingroup/Loyalty, Fairness, Purity, Authority and Care/Harm, not an internal or stable model trait.

The study tests gpt-4o-mini, claude-3-5-haiku-20241022, deepseek-chat and Wizard-Vicuna-30B-Uncensored, using two temperatures for each commercial model and one for Vicuna. Three experiments elicit an “inherent” response without an ideological label, explicit liberal or conservative role-play, and role-play with 28 biographies, fourteen per political label. Personas bundle four to nine attributes derived from summaries of Pew political typologies. The human comparison reuses Graham, Haidt and Nosek (2009): 2,212 US volunteers, including 1,174 liberals, 500 conservatives and 538 moderates or others; 62% were women and median age was 32. Human ideology was self-reported on a single item.

In aggregate, unlabeled LLM responses reproduce neither the historical liberal nor conservative pattern and tend to use higher agreement scores than humans across all five foundations. All base human-model comparisons are reported as significant, but their direction does not form a coherent liberal or conservative bias. The paper itself raises an alternative explanation: Likert acquiescence or a prompt artifact rather than ideology. Explicit liberal prompting moves outputs closer to the liberal reference, although they still diverge on Ingroup, Authority and Harm. Explicit conservative prompting does not reproduce conservative respondents; its largest deviations are Ingroup (d=.835) and Harm (d=.824).

Biographies produce the strongest divergence. Against human conservatives, conservative-labelled personas reach d=2.300 on Purity and d=2.644 on Authority, the two most extreme effects in the study. The authors interpret this as possible amplification of cultural stereotypes because generated agreement exceeds the 2009 human reference. In method comparisons, baseline and explicit-conservative outputs differ significantly only on Fairness, while persona prompts broadly shift responses, especially Purity and Authority. This supports framing and biography sensitivity, not an intrinsically conservative ideology.

Persona construction prevents causal attribution to individual demographics. Attributes are correlated and presented together, while biographies also include activism, religiosity, professional success, precarity, optimism, family roles and normative gender language. Many liberal-labelled personas combine youth, urban residence, education, atheism, technology work and activism; conservative-labelled personas combine rural residence, religion, manual work, economic hardship and family. The study can observe a response to the narrative bundle, but cannot identify whether age, ethnicity, religion, class, the explicit ideology label, prompt length or an interaction caused it. The authors acknowledge that the personas are idealized and propose modular designs for future work.

The audit found reproducibility limits that must remain visible. The methods describe the human scale as 1–6, whereas the final LLM prompt requests 0–5, without documenting recoding. The paper does not report the full item list or item count, repetitions, seeds, API dates, per-condition N, invalid-response handling or statistical unit. Responses are nested by model, temperature, item and persona, yet are aggregated and compared with independent-samples t-tests without enough information to verify independence or effective N; no multiple-comparison correction is reported. The authors acknowledge inter-model variation and a lack of robust model-level statistics, so the aggregate cannot support individual rankings.

No code, prompt files, raw outputs or analysis data were found on the official page, arXiv or GitHub searches, so the figure and table cannot be recalculated. The human reference is US-based and from 2009, sixteen years before the tested systems; the paper also calls liberals three times as numerous as conservatives, although 1,174/500 is about 2.35. The faithful conclusion is narrow: for these prompts and 2025-era models, MFT outputs are prompt-sensitive, and stereotype-rich biographies can amplify Purity and Authority beyond a historical human reference. The study does not establish synthetic personality, internal ideology, accurate group representation, demographic causality or production harm.

Español

El artículo estudia cómo cambian las respuestas de cuatro LLM a ítems de juicio de Moral Foundations Theory (MFT) cuando reciben un prompt procedimental, una identidad política explícita o una biografía estereotipada. No mide personalidad: las “personas” son descripciones redactadas que agrupan edad, género, etnia, geografía, religión, educación, situación económica y valores sociales asociados por los autores con liberalismo o conservadurismo. El objeto observado es la respuesta condicionada por prompt en cinco fundamentos, Lealtad al grupo, Equidad, Pureza, Autoridad y Cuidado/Daño, no un rasgo interno o estable del modelo.

Se prueban gpt-4o-mini, claude-3-5-haiku-20241022, deepseek-chat y Wizard-Vicuna-30B-Uncensored, con dos temperaturas para cada modelo comercial y una para Vicuna. Hay tres experimentos: respuesta “inherente” sin etiqueta ideológica, role-play explícito como liberal o conservador y role-play con 28 personas, catorce por etiqueta política. Las personas combinan entre cuatro y nueve atributos tomados de resúmenes de tipologías de Pew. Como referencia humana se reutiliza Graham, Haidt y Nosek (2009): 2.212 voluntarios estadounidenses, 1.174 liberales, 500 conservadores y 538 moderados u otros, 62 % mujeres y mediana de 32 años. La ideología humana fue una autoidentificación en un único ítem.

Agregados, los LLM sin etiqueta no reproducen ni el patrón liberal ni el conservador histórico y tienden a puntuar más alto que los humanos en los cinco fundamentos. Todas las diferencias humano-modelo de la condición base se reportan significativas, pero la dirección no forma un sesgo liberal o conservador coherente. El propio artículo plantea una explicación alternativa: aquiescencia ante escalas Likert o un artefacto del prompt, no ideología. Pedir explícitamente una perspectiva liberal acerca las respuestas a la referencia liberal, aunque siguen divergiendo en Lealtad, Autoridad y Daño. Pedir una perspectiva conservadora no reproduce a los conservadores humanos; las mayores desviaciones son Lealtad (d=0,835) y Daño (d=0,824).

Las biografías producen la divergencia más fuerte. Frente a conservadores humanos, las personas conservadoras alcanzan d=2,300 en Pureza y d=2,644 en Autoridad, los dos tamaños de efecto más extremos del estudio. Los autores interpretan este patrón como posible amplificación de estereotipos culturales: el apoyo generado supera lo observado en la referencia humana de 2009. En la comparación entre métodos, la respuesta base y el role-play conservador solo difieren significativamente en Equidad, mientras que las personas alteran ampliamente las respuestas, sobre todo en Pureza y Autoridad. Esto apoya sensibilidad al encuadre y a biografías cargadas; no demuestra una ideología conservadora intrínseca.

La construcción de personas impide atribuir causalidad a datos demográficos concretos. Los atributos están correlacionados y se presentan juntos; además, las historias incluyen activismo, religiosidad, éxito profesional, precariedad, optimismo, roles de género y lenguaje normativo. Por ejemplo, muchas personas liberales combinan juventud urbana, educación, ateísmo, trabajo tecnológico y activismo, mientras las conservadoras agrupan ruralidad, religión, trabajos manuales, dificultades económicas y familia. El estudio puede detectar una respuesta al paquete narrativo, pero no decidir si la causa es edad, etnia, religión, clase, ideología explícita, longitud del prompt o cualquier interacción. Los autores reconocen que las personas son idealizadas y proponen diseños modulares futuros.

La auditoría detecta límites de reproducibilidad que el resumen debe mantener visibles. El método describe la escala humana como 1–6, pero el prompt final exige 0–5 y no explica la recodificación. No informa la lista completa ni el número de ítems, repeticiones, seeds, fechas de API, N por condición, tratamiento de respuestas inválidas o unidad estadística. Las respuestas están anidadas por modelo, temperatura, ítem y persona, pero se agregan y comparan mediante t-tests independientes sin que pueda comprobarse independencia o N efectivo; tampoco se declara corrección por comparaciones múltiples. El artículo reconoce variación entre modelos y falta de estadística robusta a ese nivel, de modo que el agregado no permite rankings individuales.

No se localizaron código, prompts como archivos, outputs crudos ni datos de análisis en la página oficial, arXiv o búsquedas de GitHub. Por ello no pueden recalcularse la figura ni la tabla. La referencia humana es estadounidense y de 2009, anterior en dieciséis años a los modelos; además, el texto la llama tres veces más liberal que conservadora, aunque 1.174/500 equivale aproximadamente a 2,35. La conclusión fiel es acotada: bajo estos prompts y modelos de 2025, las respuestas MFT son sensibles al role-play y las biografías estereotipadas pueden amplificar Pureza y Autoridad más allá de una referencia humana histórica. No demuestra personalidad sintética, ideología interna, representación fiel de grupos, causalidad demográfica ni daño en producción.

Research question

How do the responses of four LLMs to Moral Foundations Theory items differ from liberal and conservative human data from 2009, and how do they change when explicit ideological role-play is requested or when stereotyped demographic biographies are added?

Method

Three comparative experiments with four models and seven temperature configurations. Scores 0–5 are elicited for five moral foundations under base prompt, liberal/conservative identities, and 28 biographical personas constructed with attributes associated with Pew typologies. Aggregated responses are contrasted with historical human data using independent t-tests, mean differences, and standardized effect sizes. The audit reviews the complete official PDF, the arXiv appendix, the statistical report, and the availability of artifacts.

Sample: The human reference contains 2,212 volunteers from the USA: 1,174 liberals, 500 conservatives, and 538 moderates or others; 62% women, 38% men, and a median of 32 years. The model sample comprises four LLMs with seven temperature configurations, three condition families, and 28 personas. The article does not report the total number of items, replicates, observations, or N per contrast, so the effective LLM sample size cannot be reconstructed.

Findings

  • Aggregated base responses differ from human liberals and conservatives on all five foundations.
  • LLMs tend to use higher agreement scores than humans.
  • No coherent base direction appears that reproduces the liberal or conservative human political pattern.
  • The article itself admits that the base difference may be Likert acquiescence or a prompt artifact.
  • Liberal role-play approaches the liberal reference, but continues to diverge on Loyalty, Authority, and Harm.
  • Conservative role-play does not reproduce the conservative reference.
  • The largest explicit conservative deviations are Loyalty d=0.835 and Harm d=0.824.
  • Liberal and conservative personas diverge from their human references, with more intense effects among the conservative ones.
  • Conservative personas reach d=2.300 on Purity compared to human conservatives.
  • Conservative personas reach d=2.644 on Authority, the largest reported effect.
  • The base condition and the explicitly conservative one differ significantly only on Fairness in the aggregated RQ4 comparison.
  • The authors interpret the Purity and Authority extremes as possible stereotypical amplification.
  • Personas change responses more than simple ideological labels.
  • The results support sensitivity to the prompt and to the narrative package, not to personality or internal ideology.

Limitations

  • Personality is not measured with any validated inventory or construct.
  • The personas are idealized biographies written by the authors.
  • Each persona groups multiple correlated attributes and does not allow isolating causal effects.
  • The narratives add values and normative signals beyond demographic data.
  • The political label is also provided explicitly in the persona condition.
  • The human scale is described as 1–6 and the LLM prompt uses 0–5 without documented recoding.
  • The complete list and number of administered items are not reported.
  • Replicates, seeds, API dates, or exact versions are not reported beyond a few aliases.
  • N per condition, foundation, model, or contrast is not reported.
  • The statistical unit is not clearly defined and independence cannot be verified.
  • Observations are potentially nested by model, temperature, item, and persona.
  • Heterogeneous models are aggregated although recognized inter-model variation exists.
  • Robust statistics per model are not reported.
  • No correction for multiple testing is declared.
  • The human reference is from the United States and dates from 2009.
  • The human sample has unbalanced political groups and relies on single-item self-identification.
  • The 3:1 proportion claim does not match 1,174/500, approximately 2.35:1.
  • No code, raw outputs, or public analysis data were found.
  • Tables or figures cannot be numerically reproduced.
  • Only abstract MFT items are examined, not natural decisions or conversations.
  • Results depend on iteratively optimized prompts for compliance.
  • Generalization to current models, languages, and cultures is not evaluated.

What the study does not establish

  • It does not demonstrate an intrinsic or stable ideology in any LLM.
  • It does not demonstrate synthetic personality.
  • It does not validate the 28 personas as human types.
  • It does not demonstrate human moral reasoning in the models.
  • It does not demonstrate causality of age, gender, ethnicity, religion, geography, or education.
  • It does not demonstrate accurate representation of real liberals or conservatives.
  • It does not offer reliable rankings among individual models.
  • It does not separate the effect of prompt length, ideological label, and biographical content.
  • It does not generalize to populations outside the historical US reference.
  • It does not measure behavior, persuasion, discrimination, or harm toward users.
  • It does not estimate the behavior of deployed systems.
  • It does not allow complete statistical reproduction.

Traceability

Scope: Full text

Version: AIES 2025 proceedings version, pp. 2419-2430; extended arXiv:2510.13902v1 appendices inspected. Twelve-page publisher PDF fully rendered and visually inspected; public artifact search found no code or data.

Consulted source: https://ojs.aaai.org/index.php/AIES/article/view/36727

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

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI gpt-4o-mini at temperatures 1.0 and 2.0
  • Anthropic claude-3-5-haiku-20241022 at temperatures 0.5 and 1.0
  • DeepSeek deepseek-chat at temperatures 1.0 and 1.5
  • Wizard-Vicuna-30B-Uncensored at temperature 0.7

Instruments and metrics

  • Moral Foundations Theory moral judgment items
  • Ingroup/Loyalty foundation
  • Fairness/Reciprocity foundation
  • Purity/Sanctity foundation
  • Authority/Respect foundation
  • Care/Harm foundation
  • Six-point Likert response scale described inconsistently as 1-6 and 0-5
  • Explicit liberal/conservative identity prompts
  • Twenty-eight authored biographical personas
  • Independent-samples t-tests and standardized mean differences

Data used

  • Graham, Haidt and Nosek 2009 human MFT responses
  • Pew Research Center American Trends Panel political typology summaries from 2021 and 2024
  • No public raw LLM response or analysis dataset found
  • Extended persona list in arXiv:2510.13902v1

Evidence and location

  • Research questions, contributions, and scope: AIES 2025 proceedings pp. 2419-2421, Introduction and Contribution
  • Human sample and MFT instrument: AIES 2025 proceedings pp. 2421-2422, Background and Instrument Details
  • Models, temperatures, scale, and final prompt: AIES 2025 proceedings pp. 2422-2424, Table 1 and Methodology
  • Persona construction and content: AIES 2025 proceedings pp. 2424-2425 and arXiv:2510.13902v1 Appendices A-B
  • Results and effect sizes: AIES 2025 proceedings pp. 2425-2427, Figure 1 and Table 2
  • Interpretation of stereotypes, limitations, and conclusion: AIES 2025 proceedings pp. 2427-2428, Results, Limitations and Conclusion
  • Official record and publication metadata: Official AIES article page and DOI 10.1609/aies.v8i3.36727
  • Availability of code, data, and reproducibility: Official AIES, arXiv and GitHub artifact search audited 16 July 2026
  • Complete validity report: reports/verification/article-214-mft-political-personas-validity-audit.json