The Personality Trap: How LLMs Embed Bias When Generating Human-Like Personas

Applications, bias, and safety2026arXivApproved editorial review

Authors: Jacopo Amidei, Gregorio Ferreira, Mario Muñoz Serrano, Rubén Nieto, Andreas Kaltenbrunner

Keywords: Large Language Models, Personality, Bias, Persona, AI Safety

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

5
Authors
10
Findings
38
Limitations
15
Evidence

Editorial summary

English

The study examines which demographics five LLMs invent when given only EPQR-A questions and answers and asked to write a biography with age, gender, sexual orientation, race, religion, occupation, politics, and location. The generators are GPT-3.5, GPT-4o, Claude 3.5 Sonnet, Llama 3.2-3B, and Llama 3.1-70B. The starting point is not a human sample: it consists of 826 EPQR-A response sets previously simulated by gpt-4o-2024-05-13 at temperature 1, themselves based on the sex and age of 826 Spanish university students, 655 women and 171 men, with mean ages near 19. Only answered items, not demographic labels, are passed to the new model.

For each model the authors generate ten Base populations and five MaxN and MaxP populations. MaxN changes answers to maximize Neuroticism; MaxP does the same for Psychoticism. This is 20 populations × 826 profiles × 5 models, or 82,600 persona narratives before later assessments. A single “representative” population per condition and model, selected because trial variation was considered small, but without a published selection rule, retakes EPQR-A; Base populations also answer BFI-44. The study compares scores, accuracy, MAE/RMSE, EPQR-A–BFI correlations, and Cronbach's alpha.

Base distributions show strong but model-specific stereotypes. GPT-3.5 generates 90.82% women; Claude 87.84%; Llama 3.2-3B 100% men; Llama 3.1-70B 77.94% men; GPT-4o produces 25.71% women, 44.75% men, and 29.18% non-binary personas. GPT-4o produces 29.93% LGBTQ+ personas, whereas GPT-3.5 and Llama 3.2-3B are approximately zero and Claude is 3.22%. Racially, GPT-4o is 98.98% White, GPT-3.5 88.59%, Llama 3.2-3B 100%, Llama 3.1-70B 99.08%, and Claude 62.03% White/34.82% Asian. Ages cluster at 28–32 rather than near the source sample's age of 19; locations are major US cities except London and occupations tend to require higher education.

The broad “WEIRD” label reasonably describes the Western, urban, educated, and racially narrow defaults in much of the output, but there is no complete target population against which representativeness is estimated. Only sex and age are known for the source human sample, and the 826 input profiles are synthetic responses rather than observations from that population. There are no real reference distributions for race, religion, politics, location, or sexual orientation. The percentages therefore demonstrate model defaults and generative associations, not statistical bias calculated against a representative target. The paper acknowledges in Limitations that human-derived questionnaires and comparisons with other generation methods are needed.

The most important safety result occurs under MaxP and is very large for two models. GPT-4o moves from 29.18% to 88.76% non-binary and from 29.93% to 94.12% LGBTQ+; Claude moves from 3.22% to 96.66% non-binary and from 3.22% to 99.08% LGBTQ+. Progressive labels also rise sharply. The pattern is not universal: GPT-3.5 remains at 0.31% for both categories, Llama 3.2-3B at 0%, and Llama 3.1-70B at 0.38% non-binary/3.00% LGBTQ+. The paper correctly states that this does not imply any real-world association and frames it as possible stereotyping or pathologizing inference.

The mechanism does not isolate a latent psychological trait. The prompt supplies all 24 questions and answers and asks the model to infer a biography. Maximizing P changes six answers whose content directly concerns rules, drugs, marriage, cheating, exploitation, and conventional conduct; examples show the biography paraphrasing those items. The MaxP contrast therefore activates both a psychometric label and explicit semantic cues that the model associates with identity and politics. The finding demonstrates a dangerous generator association under that textual pattern, but not that human “Psychoticism” causes or predicts gender or orientation. EPQR-A P is also not clinical psychosis and has low reliability in several reported results.

The main tables and prose do not always use the same result snapshot. The prose assigns GPT-4o 29.27±1.70% non-binary versus 29.18±1.76 in Table 1, and 29.96±1.52% LGBTQ+ versus 29.93±1.63. It gives Claude 35.16±1.72% Asian versus 34.82±1.43. More materially, it reports conservative shares of 42.68% and 33.16% for Llama 3.2-3B and Llama 3.1-70B, while Table 2 gives 32.56% and 42.93%. These divergences make the exact run supporting the narrative unclear.

Demographic difference tests compare ten Base means and five MaxN/MaxP means through two-sided t-tests. Proportions are compositional, the same 826 inputs are reused across conditions, and many categories, models, and contrasts are tested; there is no documented multiplicity correction, dependence treatment, seed, generator temperature, or assumption diagnostic. Low between-trial standard deviation can indicate a consistent stereotyped default rather than representativeness. Aggregation also changes interpretation: “non-binary” includes gender-fluid and non-conforming; “LGBTQ+” combines gay, lesbian, bisexual, pansexual, queer, asexual, and demisexual; “Centre” includes independent/moderate; “Progressive” uses a broad synonym set; and distinct religions are collapsed into “Other.”

Apparent personality fidelity is largely a closed loop. A narrative is generated from the same 24 items, then the same model answers those items again while role-playing the narrative. The paper explicitly shows semantic copies: “rarely stays in the background” mirrors item 15 and “always practicing what they preach” mirrors item 24. GPT-4o reaches Base accuracy of 97.68% for E, 93.04% for N, 98.20% for P, and 99.23% for L. This is evidence of textual preservation from the prompt, not a deep psychological construct.

BFI provides some vocabulary-level generalization: EPQR-A/BFI correlations are 0.94–0.99 for E and 0.70–0.93 for N. Yet the same model creates both narrative and answers, dimensions are explicitly verbalized, and large cross-loadings appear, EPQR-P with BFI-O is 0.50–0.71 and EPQR-E with BFI-A reaches 0.86, without formal comparison to human coefficients. The origin of the BFI “Input” row is also unexplained even though the stated input is EPQR-A responses. High alpha can reflect determinism and redundancy; P falls to 0.18–0.68 under MaxP, L for Llama 3.2-3B to 0.01–0.27, and Claude MaxN N is undefined because there is no variance. The random control deliberately destroys inter-item covariance by sampling item marginals independently, so its near-zero alpha is not a strong neutral baseline for “coherent personality.”

The public reproducibility artifact is unavailable at audit time. The paper links anonymous.4open.science/r/the_personality_trap-F487/README.md, but the server returns HTTP 410 and no equivalent public repository was found. Some model names and costs remain documented, but code, generated data, seeds, decoding parameters, logs, the “representative” sample rule, exact tests, multiplicity handling, and executable Claude/Llama versions are missing. The official record confirms only a 26-page arXiv v1 preprint, not a venue or peer review.

The work convincingly identifies a representational risk: some models connect a MaxP answer pattern to marginalized identities at extreme rates. This is a useful warning for persona-generator design and auditing. It does not validate synthetic populations as human substitutes, demonstrate real psychological correlations, or show that models “reproduce” human demography without an appropriate reference.

Español

El estudio analiza qué demografía inventan cinco LLM cuando reciben únicamente preguntas y respuestas EPQR-A y se les pide escribir una biografía con edad, género, orientación sexual, raza, religión, ocupación, política y ubicación. Los modelos generadores son GPT-3.5, GPT-4o, Claude 3.5 Sonnet, Llama 3.2-3B y Llama 3.1-70B. El punto de partida no es una muestra humana: son 826 respuestas EPQR-A previamente simuladas por gpt-4o-2024-05-13 a temperatura 1, a su vez basadas en sexo y edad de 826 universitarios españoles, 655 mujeres y 171 hombres, con medias de edad próximas a 19 años. De esa procedencia solo se pasan al nuevo modelo los ítems contestados, no las etiquetas demográficas.

Para cada modelo se generan diez poblaciones Base y cinco poblaciones MaxN y MaxP. MaxN cambia respuestas hasta llevar Neuroticismo al máximo; MaxP hace lo mismo con Psychoticism. Son 20 poblaciones × 826 perfiles × 5 modelos, es decir, 82.600 relatos-persona antes de las evaluaciones posteriores. Una sola población “representativa” por condición y modelo, elegida porque la variación entre trials se considera pequeña pero sin regla de selección publicada, vuelve a contestar EPQR-A; las Base responden también BFI-44. El trabajo compara scores, accuracy, MAE/RMSE, correlaciones EPQR-A–BFI y alfa de Cronbach.

Las distribuciones Base muestran estereotipos fuertes, pero distintos por modelo. GPT-3.5 genera 90,82 % mujeres; Claude, 87,84 %; Llama 3.2-3B, 100 % hombres; Llama 3.1-70B, 77,94 % hombres; GPT-4o reparte 25,71 % mujeres, 44,75 % hombres y 29,18 % no binario. GPT-4o produce 29,93 % LGBTQ+, mientras GPT-3.5 y Llama 3.2-3B quedan aproximadamente en cero y Claude en 3,22 %. En raza, GPT-4o es 98,98 % blanco, GPT-3.5 88,59 %, Llama 3.2-3B 100 %, Llama 3.1-70B 99,08 % y Claude 62,03 % blanco/34,82 % asiático. Las edades se concentran entre 28 y 32, no alrededor de los 19 años de la muestra fuente; las ubicaciones son grandes ciudades estadounidenses salvo Londres y los empleos suelen requerir alta educación.

La etiqueta general “WEIRD” describe bien el sesgo occidental, urbano, educado y racialmente estrecho de buena parte de las salidas, pero no existe una población objetivo completa contra la que estimar representatividad. Solo se conocen sexo y edad de la muestra humana original, y los 826 perfiles de entrada son respuestas sintéticas, no observaciones de esa población. Para raza, religión, política, ubicación u orientación sexual no hay distribución real de referencia. Por ello los porcentajes prueban defaults y asociaciones generativas del modelo, no un error estadístico calculado frente a una población representativa. El paper reconoce en Limitations que falta repetir con cuestionarios humanos y comparar con otros métodos de generación.

El resultado de seguridad más importante aparece en MaxP y es muy grande en dos modelos. GPT-4o pasa de 29,18 % a 88,76 % de identidades no binarias y de 29,93 % a 94,12 % LGBTQ+; Claude pasa de 3,22 % a 96,66 % no binario y de 3,22 % a 99,08 % LGBTQ+. También suben fuertemente las etiquetas progresistas. No es un patrón universal: GPT-3.5 queda en 0,31 % para ambas categorías, Llama 3.2-3B en 0 % y Llama 3.1-70B en 0,38 % no binario/3,00 % LGBTQ+. El propio artículo afirma correctamente que esto no implica ninguna asociación real y lo presenta como posible estereotipo o inferencia patologizante.

El mecanismo no permite atribuir el cambio a un rasgo psicológico latente. El prompt entrega las 24 preguntas y sus respuestas, y pide inferir una biografía. Maximizar P cambia seis respuestas cuyo contenido habla directamente de reglas, drogas, matrimonio, engaño, aprovechamiento y conducta convencional; los ejemplos muestran que la biografía parafrasea esos ítems. Así, el contraste MaxP activa simultáneamente la etiqueta psicométrica y señales semánticas explícitas que el modelo asocia con identidades y política. El hallazgo demuestra una asociación peligrosa del generador bajo ese patrón textual, pero no que “Psychoticism” humano cause o prediga género u orientación. Además, la escala P de EPQR-A no equivale a psicosis clínica y tiene baja fiabilidad en varios resultados.

La tabla principal y la prosa no usan siempre el mismo snapshot de resultados. El texto atribuye a GPT-4o 29,27±1,70 % no binario, mientras la Tabla 1 dice 29,18±1,76; da 29,96±1,52 % LGBTQ+ frente a 29,93±1,63; para Claude cita 35,16±1,72 % asiático frente a 34,82±1,43. Más materialmente, afirma porciones conservadoras de 42,68 % y 33,16 % para Llama 3.2-3B y Llama 3.1-70B, pero la Tabla 2 informa 32,56 % y 42,93 %. Estas divergencias impiden saber qué corrida sostiene exactamente el relato.

Los tests de diferencias demográficas se aplican a diez promedios Base y cinco MaxN/MaxP mediante t-tests bilaterales. Las proporciones son composicionales, las mismas 826 entradas se reutilizan en todas las condiciones y se prueban numerosas categorías, modelos y contrastes; no se documenta corrección por multiplicidad, tratamiento de dependencia, seeds, temperatura de los cinco generadores o diagnóstico de supuestos. La baja desviación entre trials puede significar consistencia del default estereotípico, no representatividad. Las agregaciones también cambian la interpretación: “non-binary” incluye gender-fluid y non-conforming; “LGBTQ+” agrupa gay, lesbiana, bisexual, pansexual, queer, asexual y demisexual; “Centre” incluye independent/moderate; “Progressive” usa un conjunto amplio; y religiones distintas se colapsan en “Other”.

La aparente fidelidad de personalidad es en gran medida un circuito cerrado. El relato se genera desde los mismos 24 ítems y después el mismo modelo vuelve a contestarlos interpretando el relato. El paper muestra explícitamente copias semánticas: “rarely stays in the background” replica el ítem 15 y “always practicing what they preach” el ítem 24. GPT-4o alcanza accuracy Base de 97,68 % en E, 93,04 % en N, 98,20 % en P y 99,23 % en L. Es evidencia de preservación textual del prompt, no de un constructo psicológico profundo.

La BFI aporta cierta generalización de vocabulario: las correlaciones EPQR-A/BFI en E son 0,94–0,99 y en N 0,70–0,93. Sin embargo, el mismo modelo genera relato y respuestas, las dimensiones se verbalizan de forma explícita y existen correlaciones cruzadas muy altas, por ejemplo EPQR-P con BFI-O de 0,50–0,71 y EPQR-E con BFI-A de hasta 0,86, que no se contrastan formalmente con coeficientes humanos. La procedencia de la fila BFI “Input” tampoco se explica, pese a que el método define la entrada como respuestas EPQR-A. Las alfas altas pueden reflejar determinismo y redundancia; P es baja, entre 0,18 y 0,68 en MaxP, L de Llama 3.2-3B cae a 0,01–0,27 y el N de Claude MaxN es indefinido por ausencia de varianza. El control aleatorio destruye deliberadamente la covarianza entre ítems al muestrearlos por marginales, de modo que su alfa cercana a cero no es un baseline neutral fuerte para “personalidad coherente”.

La reproducibilidad pública no está disponible en la fecha de auditoría. El paper enlaza anonymous.4open.science/r/the_personality_trap-F487/README.md, pero el servidor devuelve HTTP 410; no se localizó un repositorio público equivalente. Quedan especificados algunos nombres de modelo y costes, pero faltan código, datos generados, seeds, parámetros de decoding, logs, regla de muestra “representativa”, tests exactos, corrección de multiplicidad y versiones ejecutables completas de Claude/Llama. La superficie oficial solo confirma arXiv v1, 26 páginas, sin venue o peer review confirmado.

El trabajo identifica de forma convincente un riesgo de representación: ciertos modelos conectan un patrón de respuestas MaxP con identidades marginalizadas de manera extrema. Es una alarma útil para diseño y auditoría de generadores de personas. No valida poblaciones sintéticas como sustitutas de humanos, no demuestra correlaciones psicológicas reales y no permite afirmar que los modelos “reproducen” demografía humana sin una referencia apropiada.

Research question

What sociodemographic attributes do five LLMs invent when transforming EPQR-A responses into persons, how do they change when maximizing Neuroticism or Psychoticism, and to what extent does the narrative preserve the personality responses?

Method

826 EPQR-A responses already synthetically answered by GPT-4o are reused. Five models generate ten Base populations and five MaxN/MaxP populations of 826 persons each. Eight sociodemographic attributes are added, proportions are compared with t-tests, and one population per condition re-answers the EPQR-A; the Base populations also answer the BFI. Accuracy, MAE/RMSE, correlations, and Cronbach's alpha are computed. The audit cross-checks all tables, prompts, and appendices and verifies the status of the linked artifact.

Sample: Five generators x 826 profiles x 20 populations per model (10 Base, 5 MaxN, and 5 MaxP) produce 82,600 narratives. The psychometric evaluation retains one population per condition and model, but does not publish the rule for selecting it. The demographic t-tests operate on 10 Base trials against 5 manipulated trials; no seeds, losses, parse failures, or total calls/tokens are reported.

Findings

  • Base populations exhibit distinct and extreme defaults by model in gender, race, religion, politics, age, occupation, and location.
  • Four models generate 88.59-100% white persons; Claude generates 62.03% white and 34.82% Asian.
  • MaxP raises non-binary/LGBTQ+ to 88.76/94.12% in GPT-4o and 96.66/99.08% in Claude.
  • The MaxP increase does not generalize: GPT-3.5, Llama 3.2-3B, and Llama 3.1-70B remain near zero in those categories.
  • The observed association is a signal of generator stereotype, and the paper explicitly denies that it implies a real relationship.
  • The biographies preserve phrases from the EPQR-A items and allow very high re-administration accuracies, especially in GPT-4o and Claude.
  • EPQR-E/BFI-E correlates 0.94-0.99 and EPQR-N/BFI-N 0.70-0.93, along with substantial cross-loadings.
  • The P scale shows low reliability, especially in MaxP, and several extreme scales lose variance.
  • The prose and tables contain several incompatible figures, including the conservative proportions of both Llama models.
  • The official code artifact cited by the paper is no longer accessible.

Limitations

  • The input is a synthetic population generated by GPT-4o, not human responses.
  • The originating human framework only provides sex and age and is already a narrow university sample.
  • There is no target population to measure representativeness of race, religion, politics, location, or sexual orientation.
  • The percentages describe generator defaults, not statistical bias against a defined real distribution.
  • The prompt receives complete questions and answers, which contain direct semantic signals for the biography.
  • MaxP simultaneously changes the score and the explicit content about rules, drugs, marriage, and conduct.
  • The Psychoticism label is not isolated from the wording of its six items.
  • EPQR-A P does not equate to clinical psychosis; the terminology may favor an incorrect pathologizing reading.
  • The effect is not uniform across models and is concentrated extremely in GPT-4o and Claude.
  • The gender and orientation aggregations combine distinct identities and affect the published rates.
  • The politics, race, and religion categories are narrow and culturally situated.
  • Islam, Hinduism, and Buddhism are collapsed into Other; claims of absence are not auditable from the aggregated tables.
  • No tables for age, location, or occupation are shown despite using them for the WEIRD conclusions.
  • The prose and tables do not agree on several means and standard deviations.
  • The discrepancy in conservative percentages for Llama changes which model appears more conservative.
  • 10 Base trials are compared with 5 manipulated trials without justifying the balance.
  • The proportions sum to 100% per category, but are treated with separate t-tests.
  • The same 826 inputs are reused; the treatment of dependence is not documented.
  • There is no correction for the numerous comparisons of categories, conditions, and models.
  • No tests of assumptions, effect intervals, or a hierarchical model per entry and trial are reported.
  • Temperatures, seeds, and decoding parameters for the five evaluated generators are missing.
  • The published identifiers for GPT and Claude are not all unambiguously executable snapshots.
  • A single "representative" population is chosen for RQ3 without a reproducible criterion.
  • Uncertainty across trials is not propagated to accuracy, correlations, or alpha.
  • The same model generates the narrative and answers the questionnaires from that narrative.
  • EPQR-A is re-administered after having literally exposed its items and answers.
  • High accuracy may measure copying or paraphrase memory, not underlying personality.
  • BFI shares trait vocabulary and does not fully break the self-consistency circuit.
  • Correlations are compared narratively, not through equivalence with human coefficients.
  • The BFI Input row has no clear methodological provenance.
  • High alpha does not establish validity; it may result from deterministic or redundant responses.
  • The P scale shows alpha 0.18-0.68 in MaxP, and the N of Claude MaxN is undefined.
  • The random baseline preserves marginals but destroys by design the covariance among items.
  • Behavior, decision-making, interaction, predictive validity, or human judgment of authenticity are not evaluated.
  • Languages, cultures, alternative prompts, or robustness of the categories are not studied.
  • The official code link returns HTTP 410 and no public mirror was found.
  • Without code and data, normalization, tests, parsing, and sample selection cannot be reconstructed.
  • The study is arXiv v1 and does not credit a venue or peer review.

What the study does not establish

  • It does not establish that the generated populations are representative of a defined human population.
  • It does not demonstrate that personality and demographics maintain the same correlations as in humans.
  • It does not imply a real relationship between Psychoticism, non-binary identity, or LGBTQ+ orientation.
  • It does not demonstrate causality of a latent trait because it directly changes the textual content of the items.
  • It does not validate EPQR-A or BFI as equivalent psychological measurements in LLMs.
  • It does not prove that high narrative self-consistency corresponds to a deep or stable personality.
  • It does not justify using synthetic persons as substitutes for participants, surveys, or user research.
  • It does not allow faithful reproduction with the public artifact available on the audit date.
  • It does not confirm generalization beyond English, five models, and a single prompt.
  • It does not confirm peer-reviewed publication.

Traceability

Scope: Full text

Version: arXiv:2602.03334v1, submitted 3 February 2026; preprint, 26 pages

Consulted source: https://arxiv.org/pdf/2602.03334v1

Review: Codex full-text, bilingual-fidelity, 26-page visual, arXiv-v1, synthetic-input, demographic-baseline, MaxN-MaxP, protected-identity, psychometric-circularity, table-prose-consistency, compositional-statistics, aggregation, construct-validity and artifact-availability audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • gpt-4o-2024-05-13 at temperature 1, generator of the 826 input questionnaires
  • GPT-3.5, reported as gpt3.5-turbo-0125
  • GPT-4o, reported as gpt4o-2024-11-20
  • Claude 3.5 Sonnet, exact snapshot incompletely specified
  • Llama 3.2-3B on Amazon AWS, exact checkpoint and serving configuration unspecified
  • Llama 3.1-70B on Amazon AWS, exact checkpoint and serving configuration unspecified

Instruments and metrics

  • Eysenck Personality Questionnaire Revised-Abbreviated (EPQR-A), 24 yes/no items
  • EPQR-A Extraversion, Neuroticism, Psychoticism and Lie scales, six items each
  • Big Five Inventory (BFI), 44 Likert items
  • MaxN and MaxP item-answer interventions
  • Structured JSON persona and eight sociodemographic attributes
  • Post-hoc canonical demographic category aggregation
  • Two-sided t-tests
  • Accuracy, precision, recall, specificity, MAE and RMSE
  • Pearson cross-instrument correlations
  • Cronbach's alpha

Data used

  • 826 synthetic English EPQR-A response sets from Ferreira et al. (2025)
  • Source demographic frame: 655 women and 171 men from a Spanish university sample, mean age about 19
  • Ten Base and five MaxN/MaxP populations per generator
  • 82,600 generated persona narratives by arithmetic reconstruction
  • One selected population per model and condition for questionnaire re-administration
  • Anonymous 4open.science artifact F487, unavailable with HTTP 410 at audit time

Evidence and location

  • Metadata, authors, abstract, version, and status: Official arXiv:2602.03334v1 surface, submitted 3 February 2026
  • Provenance of 826 inputs, models, and costs: Paper, pp. 4-5, Section 3.1 and footnotes 1-3
  • Trials, selection of one population, and metrics: Paper, pp. 5-6, Sections 3.2-3.3
  • Base and MaxN/MaxP distributions: Paper, pp. 6-8, Tables 1-3
  • Divergences between prose and tables: Paper, pp. 6-8, narrative after Tables 1-3 compared with table cells
  • EPQR-A scores and difference tests: Paper, p. 9, Table 4
  • Accuracy, MAE, and RMSE: Paper, pp. 25-26, Appendix D, Tables A4-A5
  • EPQR-A-BFI correlations: Paper, p. 10, Tables 5-6
  • Reliability and random baseline: Paper, p. 11, Tables 7-8
  • Interpretation, caution, and acknowledged limitations: Paper, pp. 11-14, Sections 5-7
  • Semantic copying and MaxP examples: Paper, pp. 9 and 18-20, Section 4.2 and Appendix A
  • Generation and re-administration prompts: Paper, pp. 21-23, Appendix B
  • Demographic aggregation rules: Paper, pp. 23-24, Appendix C
  • Comprehensive visual inspection: Paper, all 26 rendered pages, including every table, figure, prompt, and appendix page
  • Non-reproducible artifact: Paper-linked anonymous.4open.science artifact F487 returned HTTP 410 on 15 July 2026; public repository search found no equivalent