Do Psychometric Tests Work for Large Language Models? Evaluation of Tests on Sexism, Racism, and Morality

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Jana Jung, Marlene Lutz, Indira Sen, Markus Strohmaier

Keywords: Large Language Models, Personality, Psychometrics, Personality Control, AI Safety

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

4
Authors
9
Findings
30
Limitations
9
Evidence

Editorial summary

English

This study asks whether three human questionnaires, the 22-item Ambivalent Sexism Inventory (ASI), 8-item Symbolic Racism 2000 Scale (SR2K), and 30-item Moral Foundations Questionnaire (MFQ), yield reliable and valid LLM scores. It evaluates 17 snapshots: Centaur; four Gemma 3 sizes; Llama 3.1 8B and 70B and Llama 3.3 70B; Mistral 7B and Mistral Large; five Qwen 2.5/3 variants; and Gemini 2.5 Flash/Pro. Each item is administered separately in a constrained-answer prompt and repeated under five seeds at each model's default temperature. Reliability here is not alpha, factor reliability, or temporal stability. It is the proportion of exactly unchanged item answers between the original and three perturbations: a GPT-5-generated alternate form reviewed by two authors, reversed answer-option order, and changing the prompt terminator from a colon to a question mark. As a baseline only for alternate forms, 150 U.S. Prolific participants answer originals and paraphrases in the same session; six fail attention checks, leaving 144. Answers are often stable under punctuation and alternate wording, but several models change sharply when options are reversed. The reversed-order table reaches 0.00 on SR2K for Gemma 3 1B, 0.19 on ASI for Qwen 2.5 7B, and about 0.27–0.28 on ASI/SR2K for Llama 3.1 8B. Describing the overall pattern as moderate reliability is defensible only for these specific perturbations. Convergent validity uses three model-level Spearman correlations and obtains the selected theoretical directions: racism–sexism rho=0.47, fairness–hostile sexism rho=−0.37, and authority–benevolent sexism rho=0.43. These are moderate associations from 17 points, without reported p-values, intervals, multiplicity correction, a full nomological network, or discriminant-validity analysis. They support relative coherence among three selected scale relations, not confirmation of the constructs. The paper's most important contribution is ecological validity, comparing questionnaire ranks with seven behavioral proxies. Sexism uses recommendation letters for 24 male and 24 female profiles and averages five dictionary-based odds ratios. Racism uses housing recommendations for paired Black/white profiles in ten cities and the difference in neighborhood opportunity index. Morality uses 227 dilemmas and measures whether advice matches an authority, care, fairness, ingroup, or purity-aligned action as classified by GPT-4o. Centaur fails downstream instructions and is excluded, so ecological correlations use 16 models: sexism −0.24; racism −0.62; authority −0.13; care −0.10; fairness 0.21; ingroup −0.12; purity 0.21. None shows the clear positive association needed to treat a questionnaire score as a predictor of its corresponding proxy; five are negative and two weakly positive. The strongest result is the racism reversal: models appearing less racist on SR2K tend to produce more unequal housing recommendations. This supports the warning against treating human questionnaire scores as behaviorally validated LLM properties. The title is broader than the design, however. Downstream tasks remain heterogeneous constructed proxies, not general behavior or deployment outcomes. The statistical unit is a small set of 16 family-related models rather than independent draws; no p-values, intervals, or adjustment across seven comparisons are shown. Moral behavior is itself judged by GPT-4o. Two authors annotate 100 answers and the judge agrees 88%, without kappa, classwise error, or uncertainty propagation. Missing responses also matter because test scores average available answers. Public original SR2K diagnostics contain 15 missing of 40 trials for Llama 3.1 8B, 11/40 for Gemini Flash, and 16/40 for Gemini Pro, while figures do not disclose effective denominators. Reproducibility is substantial but incomplete. The repository releases items, prompts, outputs, notebooks, and core utilities, but Table 5 mistakenly maps Llama 3.1 70B to the 8B model ID; JSON outputs preserve only a filename label, not the actual model_id, so they cannot resolve which checkpoint ran. In addition, validity_eval.py requires absent rel_reversed.json and housing_per_model.csv files, and the housing aggregation code is not supplied. The faithful conclusion is bounded: for these three tests, model snapshots, and proxies, format stability is uneven and questionnaire model rankings do not predict behavioral rankings. The study does not prove every psychometric test useless for every LLM, does not measure persistent personality, and does not make its proxies ground truth. It shows that a score should not be interpreted without use-specific validation of reliability, missing-response handling, and behavioral correspondence.

Español

Este trabajo pregunta si tres cuestionarios humanos, Ambivalent Sexism Inventory (ASI, 22 ítems), Symbolic Racism 2000 Scale (SR2K, 8) y Moral Foundations Questionnaire (MFQ, 30), producen en LLM puntuaciones fiables y válidas. Evalúa 17 snapshots: Centaur; cuatro Gemma 3; Llama 3.1 8B y 70B y Llama 3.3 70B; Mistral 7B y Mistral Large; cinco Qwen 2.5/3; y Gemini 2.5 Flash/Pro. Cada ítem se presenta por separado con respuesta cerrada y se repite con cinco semillas usando la temperatura predeterminada de cada modelo. La fiabilidad no significa alfa, estructura factorial ni estabilidad temporal: es el porcentaje de respuestas exactamente iguales entre el original y tres perturbaciones, una forma alternativa generada por GPT-5 y revisada por dos autores, el orden invertido de opciones y el cambio de dos puntos por interrogación al final del prompt. Como referencia solo para la forma alternativa, 150 participantes estadounidenses de Prolific contestan original y paráfrasis en la misma sesión; se excluyen seis por atención y quedan 144. La puntuación suele ser estable ante puntuación y forma alternativa, pero varios modelos cambian mucho al invertir opciones. La tabla de orden invertido llega a 0,00 en SR2K para Gemma 3 1B, 0,19 en ASI para Qwen 2.5 7B y aproximadamente 0,27–0,28 para Llama 3.1 8B en ASI/SR2K. Llamar al conjunto “fiabilidad moderada” es razonable solo para estas perturbaciones concretas. La validez convergente usa tres correlaciones de Spearman entre medias por modelo y obtiene las direcciones teóricas seleccionadas: racismo–sexismo rho=0,47, fairness–sexismo hostil rho=−0,37 y authority–sexismo benevolente rho=0,43. Son asociaciones moderadas de 17 puntos, sin p, intervalos, corrección múltiple, red nomológica completa ni prueba discriminante; por tanto apoyan coherencia relativa entre tres escalas, no confirman por sí solas los constructos. La contribución más importante es la validez ecológica: compara el ranking del test con siete proxies de conducta. Sexismo usa cartas de recomendación para 24 perfiles masculinos y 24 femeninos y promedia cinco odds ratios de vocabulario; racismo usa recomendaciones de vivienda para perfiles Black/white en diez ciudades y la diferencia del índice de oportunidad; moralidad usa 227 dilemas y mide si el consejo coincide con una acción alineada con authority, care, fairness, ingroup o purity, clasificada por GPT-4o. Centaur no sigue las instrucciones downstream y se excluye, de modo que las correlaciones ecológicas usan 16 modelos: sexismo −0,24; racismo −0,62; authority −0,13; care −0,10; fairness 0,21; ingroup −0,12; purity 0,21. Ninguna muestra la asociación positiva clara que haría falta para interpretar la puntuación del cuestionario como predictor de ese proxy; cinco son negativas y dos débilmente positivas. El resultado más fuerte es el contrasentido en racismo: los modelos que parecen menos racistas en SR2K tienden a producir recomendaciones de vivienda más desiguales. Esto sí respalda la advertencia de no usar directamente puntuaciones humanas como propiedades conductualmente validadas de un LLM. El título, sin embargo, es más amplio que el diseño. Las tareas downstream siguen siendo proxies construidos y distintos entre sí, no conducta general ni outcomes de despliegue. La unidad estadística es un conjunto pequeño de 16 modelos relacionados por familia, no ejecuciones independientes; no se muestran p, intervalos ni ajuste para siete comparaciones. El juicio moral depende de GPT-4o: dos autores anotan 100 respuestas y el juez coincide en 88%, sin kappa, error por clase o propagación de esa incertidumbre. Hay no-respuestas relevantes que la puntuación omite al promediar disponibles: en los datos públicos del SR2K original faltan 15 de 40 respuestas para Llama 3.1 8B, 11/40 para Gemini Flash y 16/40 para Gemini Pro, sin que las figuras muestren el denominador efectivo. La reproducibilidad es sustancial pero incompleta. El repositorio aporta ítems, prompts, salidas, notebooks y utilidades, pero la tabla asigna por error a Llama 3.1 70B el ID de 8B; los JSON solo conservan el nombre de archivo, no el model_id real, así que no resuelven qué checkpoint produjo esas salidas. Además, validity_eval.py requiere rel_reversed.json y housing_per_model.csv ausentes, y no aparece el código que construye el agregado de vivienda. La conclusión fiel es acotada: en estos tres tests, estos snapshots y estos proxies, la estabilidad ante formato es desigual y el orden relativo de los modelos en cuestionarios no predice el orden en conducta. El estudio no prueba que todo test psicométrico sea inútil para todo LLM, no mide personalidad persistente y no convierte sus proxies en verdad de terreno; demuestra que una puntuación no debe interpretarse sin validar fiabilidad, manejo de no-respuestas y correspondencia conductual para el uso concreto.

Research question

Are the ASI, SR2K, and MFQ scores obtained from 17 LLMs reliable across item and prompt variations, coherent across scales, and ecologically valid with respect to behavioral tasks?

Method

22 ASI items, 8 SR2K items, and 30 MFQ items are administered individually to 17 models, with five seeds, original form, paraphrase, inverted option order, and changed prompt ending. Exact response agreement is measured; 144 United States participants serve as a reference for original versus alternative form. Convergent validity uses three Spearman correlations between tests. Ecological validity correlates test rankings and seven downstream proxies of letters, housing, and moral advice in 16 models, after excluding Centaur.

Sample: Seventeen LLMs with five seeds per condition. Ecological validity uses 16 because Centaur did not sufficiently follow the downstream instructions. The human benchmark recruits 150 adults from the United States on Prolific with age and education quotas; six are excluded for attention checks and 144 are analyzed, who answer the original and alternative form in the same session.

Findings

  • The current source is arXiv:2510.11254v2, checked on 27 January 2026, with 31 pages and EACL Main 2026 acceptance declared in the repository.
  • The 31 pages were rendered and visually inspected; the current PDF matches byte for byte with the cache and has SHA-256 c3a71f464a8cdc5223a9dcb237a79021cbbb46af63678fbd1dda1121513ed38e.
  • Consistency is usually higher when only scoring is changed and lower when the order of options is inverted.
  • Reported convergent correlations: 0.47 racism-sexism, -0.37 fairness-hostile sexism, and 0.43 authority-benevolent sexism.
  • Ecological correlations: -0.24 sexism, -0.62 racism, -0.13 authority, -0.10 care, 0.21 fairness, -0.12 ingroup, and 0.21 purity.
  • Five of seven ecological associations have a negative direction and the two positive ones are weak.
  • The racism association is a moderate-strong inversion: better SR2K ranking coincides with worse housing proxy.
  • The public data contain substantial non-responses in SR2K for several models, omitted when averaging available scores.
  • The repository publishes a large part of the inputs, outputs, and analyses, but does not run end-to-end validity with the versioned files.

Limitations

  • Only three questionnaires, three normative domains, English prompts, and a selection of 2025 models are studied.
  • Reliability is operationalized as exact agreement under three perturbations, not as internal consistency, factorial structure, invariance, or temporal test-retest.
  • The alternative forms are generated by GPT-5 and reviewed by two authors, not psychometrically validated with an independent human study.
  • The human benchmark covers original versus alternative form, not inverted order or scoring.
  • Original and alternative form are administered in the same human session, with possible recall.
  • Only three convergent correlations chosen by theory are tested; there is no complete matrix, discriminant, or factorial analysis.
  • Convergence conclusions are based on 17 means per model.
  • Ecological correlations are based on 16 rankings per model after excluding Centaur.
  • Several points come from sizes of the same family and are not independent observations of an LLM population.
  • No p-values, confidence intervals, power, or correction for the seven ecological tests are reported.
  • Averaging five seeds before correlating hides within-model uncertainty.
  • Default temperatures differ between models, so model and sampling policy are partially confounded.
  • The letters task measures word dictionaries, not real hiring decisions or all modes of gender bias.
  • The housing task restricts the model to constructed sets of 20 neighborhoods and a composite index, not real housing outcomes.
  • Sex is limited to binary categories and racism to Black/white profiles in United States cities.
  • The five moral dimensions do not use a homogeneous source: three come from DailyDilemmas and two from Reddit.
  • GPT-4o judges the moral behavior of other LLMs, introducing error and possible shared bias.
  • The 88% judge agreement with 100 human annotations does not include kappa, class sensitivity, or interval.
  • Direct tests may activate guardrails different from downstream prompts; the study proposes this as an explanation but does not identify it causally.
  • Non-responses are omitted from the means and effective denominators per score are not shown.
  • In original SR2K, 15/40 responses are missing for Llama 3.1 8B, 11/40 for Gemini Flash, and 16/40 for Gemini Pro in the public artifact.
  • The ID table assigns Llama 3.1 70B the checkpoint meta-llama/Llama-3.1-8B-Instruct.
  • The JSON files named 70B do not store model_id or revision and do not resolve whether the error is only in the table or also in the execution.
  • Immutable Hugging Face revisions or exact Gemini versions are not fixed.
  • The repository has no immutable release or tag for the paper.
  • validity_eval.py depends on results/rel_reversed.json and housing_per_model.csv, absent in the audited commit.
  • The code that derives the housing aggregates from the published outputs is not located.
  • The environment is an extensive Linux-specific Conda export, not a minimal reproducible lockfile.
  • There is no CI or automated test for reproducing tables and figures.
  • The work does not replicate results in other languages, cultures, model versions, or test families.

What the study does not establish

  • It does not establish that all human psychometric tests fail for all LLMs.
  • It does not establish that ASI, SR2K, or MFQ lack descriptive utility in any design.
  • It does not demonstrate persistent psychological traits or traits equivalent to humans in the models.
  • It does not turn letters, housing, or moral advice into complete or ground-truth measures of the construct.
  • It does not causally demonstrate that guardrails or format explain the discrepancies.
  • It does not demonstrate significance or population stability of each ecological correlation.
  • It does not generalize to models, languages, prompts, tests, or tasks outside the sample.
  • It does not allow unambiguous verification of the nominal Llama 3.1 70B checkpoint.
  • It does not justify interpreting an individual score without specific behavioral validation.

Traceability

Scope: Full text

Version: arXiv:2510.11254v2, revised 27 January 2026, 31 pages; accepted to EACL Main 2026 according to the authors' repository

Consulted source: https://arxiv.org/abs/2510.11254

Review: Codex complete bilingual full-text fidelity pass, current arXiv v2 and byte-level PDF verification, all-page visual inspection, table and figure extraction, reliability-construct audit, ecological-proxy audit, public output missingness audit, model-identity reconciliation, and end-to-end repository reproducibility check; summaries written from the complete paper and released artifacts rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • marcelbinz/Llama-3.1-Centaur-70B
  • Gemma 3 1B, 4B, 12B and 27B instruction-tuned
  • Llama 3.1 8B and nominal 70B Instruct, plus Llama 3.3 70B Instruct
  • Mistral 7B Instruct v0.3 and Mistral Large 123B
  • Qwen 2.5 7B, 14B, 32B and 72B Instruct; Qwen 3 4B Instruct 2507
  • Gemini 2.5 Flash and Gemini 2.5 Pro
  • GPT-5 for alternate-item generation
  • GPT-4o as morality downstream-task judge

Instruments and metrics

  • Ambivalent Sexism Inventory, 22 items with hostile and benevolent sexism subscales
  • Symbolic Racism 2000 Scale, 8 items
  • Moral Foundations Questionnaire, 30 items across care, fairness, ingroup, authority, and purity
  • Exact-answer consistency under alternate form, reversed answer order, and changed prompt terminator
  • Spearman rank correlation for selected convergent and ecological relations
  • Dictionary-based gendered-language odds ratios
  • Neighborhood opportunity-index difference
  • GPT-4o classification of foundation-aligned advice

Data used

  • Original and GPT-5-generated alternate ASI, SR2K, and MFQ items
  • Five seeded runs per model, test, variation, and downstream task
  • 48 reference-letter profiles: 24 male and 24 female across age and occupation
  • Housing recommendation prompts for paired Black/white profiles across ten U.S. cities
  • 123 selected DailyDilemmas for care, fairness, and ingroup
  • 104 Reddit-derived dilemmas for authority and purity
  • Public authors' repository at commit 883bfe528b7ceb51c758015f4ac160ff404a9c22

Evidence and location

  • Version, authors, abstract, and review date: arXiv:2510.11254v2 metadata and pages 1-2 checked 15 July 2026
  • Design, models, tests, and metrics: Sections 3-4, pages 3-6
  • Main results of reliability and ecological validity: Sections 5-6 and Figures 2-3, pages 6-9
  • Human sample and detailed tables: Appendices C-G, pages 19-25
  • Convergent correlations and downstream scores: Figure 12 and Table 10, page 24; Figure 13, page 25
  • Items and alternative forms: Appendix H, pages 26-31
  • Availability, non-responses, model identity, and execution: Authors' GitHub repository commit 883bfe528b7ceb51c758015f4ac160ff404a9c22 audited 15 July 2026
  • Methodological and reproducibility audit: reports/verification/article-185-validity-and-reproducibility-audit.json
  • Complete visual inspection: All 31 PDF pages rendered and visually inspected on 15 July 2026