On the Reliability of Psychological Scales on Large Language Models

Evaluation and psychometric validity2024ACL AnthologyApproved editorial review

Authors: Jen-tse Huang, Wenxiang Jiao, Man Ho Lam, Eric John Li, Wenxuan Wang, Michael R. Lyu

Keywords: Large Language Models, Psychological assessment, Personality traits, Big Five Inventory, Personality emulation

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

6
Authors
8
Findings
9
Limitations
9
Evidence

Editorial summary

English

This EMNLP 2024 paper examines the stability of Big Five Inventory (BFI) scores when administered to LLMs under prompt changes. It constructs a 5 instruction templates × 5 item versions × 10 languages × 5 choice-label formats × 2 choice orders factorial, yielding 2,500 configurations per model. The BFI has 44 five-point items; items are shuffled and presented in random blocks of 17–27. GPT-4-Turbo creates four paraphrases per item; Google Translate and DeepL produce nine translations from English, with only a sample manually checked. At temperature zero, the study evaluates GPT-3.5-Turbo-1106, GPT-4-Turbo-1106, Gemini-1.0-Pro, and LLaMA-3.1-8B. For GPT-3.5, DBSCAN labels 77 configurations, 3.08%, as outliers, concentrated in numerical labels, descending order, and Arabic or Chinese. Only 7 of 135 one-level-versus-rest comparisons exceed an absolute difference of .15, although the table contains many small p-values because .15 is a descriptive threshold rather than a validated psychometric criterion. GPT-3.5 mean ± SD scores are 4.31±.44 for Openness, 4.15±.39 for Conscientiousness, 3.89±.43 for Extraversion, 4.13±.38 for Agreeableness, and 2.35±.42 for Neuroticism. GPT-4, Gemini, and LLaMA have distinct distributions and outlier rates of 5.6%, 4.2%, and 4.4%; LLaMA is more dispersed. Comparing GPT-3.5 dispersion with more heterogeneous human norms finds lower model variability, but this does not prove reliability: it is also compatible with temperature zero, response bias, or determinism. The test–retest study queries GPT-3.5 biweekly from September 2023 through January 2024 across snapshots 0613 and 1106. Its conclusion of no variation conflicts with Table 2: maximum Agreeableness differs from the mean with p displayed as .00 and is marked “No” for equal means. The analysis also selects extrema post hoc, treats failure to reject as acceptance of equality, and reports no test–retest coefficient. A second part tests three ways to alter GPT-3.5 self-ratings: emotional narratives, explicit assignment of ten extreme profiles, and embodiment of eight heroes and eight villains. Narratives barely move the distribution; among question answering, biography, and portray methods, only direct portray instructions clearly shift it. Extreme instructions separate every targeted dimension from default (p < .001): the largest change is minimum Extraversion at −1.71, followed by maximum Neuroticism at +1.03. Heroes stay near the default assistant profile, while villains spread more broadly. This demonstrates that self-report outputs respond to descriptions containing the target trait, not accurate representation of people or human populations: there is no target human data, behavioral validation, external criterion, or character-fidelity evaluation. The main overreach is calling aggregate robustness to perturbations “internal consistency reliability.” The paper itself acknowledges that its transformations preclude Cronbach's alpha and that reliability does not imply validity. The defensible contribution is therefore a broad map of BFI sensitivity to formats, languages, and instructed personas; it neither validates the BFI as a measure of LLM personality nor supports replacing human participants.

Español

Este trabajo de EMNLP 2024 examina la estabilidad de las puntuaciones del Big Five Inventory (BFI) cuando se administra a LLM bajo cambios de prompt. Construye un factorial de 5 plantillas de instrucción × 5 versiones de los ítems × 10 idiomas × 5 formatos de etiqueta × 2 órdenes de las opciones: 2.500 configuraciones por modelo. El BFI tiene 44 ítems en escala de cinco puntos; los ítems se barajan y se presentan en bloques aleatorios de 17–27. GPT-4-Turbo genera cuatro paráfrasis por ítem; Google Translate y DeepL producen nueve traducciones desde el inglés y solo una muestra se verifica manualmente. Con temperatura 0 se evalúan GPT-3.5-Turbo-1106, GPT-4-Turbo-1106, Gemini-1.0-Pro y LLaMA-3.1-8B. En GPT-3.5, 77 configuraciones, 3,08 %, se clasifican como outliers mediante DBSCAN; se concentran en etiquetas numéricas, orden descendente y árabe o chino. Solo 7 de 135 comparaciones de un nivel frente al resto superan una diferencia absoluta de 0,15, aunque la tabla muestra numerosos p-valores pequeños porque el umbral de 0,15 es descriptivo y no un criterio psicométrico validado. Las medias ± desviación estándar de GPT-3.5 son 4,31±0,44 en apertura, 4,15±0,39 en responsabilidad, 3,89±0,43 en extraversión, 4,13±0,38 en amabilidad y 2,35±0,42 en neuroticismo. GPT-4, Gemini y LLaMA presentan distribuciones distintas y tasas de outliers de 5,6 %, 4,2 % y 4,4 %; LLaMA aparece más disperso. Comparar la dispersión de GPT-3.5 con normas humanas más heterogéneas muestra menor variabilidad del modelo, pero eso no prueba fiabilidad: también es compatible con temperatura cero, sesgo de respuesta o determinismo. El test–retest consulta GPT-3.5 cada dos semanas entre septiembre de 2023 y enero de 2024, atravesando las versiones 0613 y 1106. La conclusión de que no hay variación contradice su Tabla 2: el máximo de Agreeableness difiere de la media con p mostrado como 0,00 y queda marcado «No» en igualdad de medias. Además, se seleccionan extremos post hoc, se interpreta no rechazar como aceptar igualdad y no se aportan coeficientes de test–retest. La segunda parte prueba tres formas de alterar la autoevaluación de GPT-3.5: narrativas emocionales, asignación explícita de diez perfiles extremos y personificación de ocho héroes y ocho villanos. Las narrativas apenas cambian la distribución; entre question answering, biography y portray, solo la instrucción directa portray la desplaza con claridad. Las instrucciones extremas separan todas las dimensiones respecto al default (p < 0,001): el mayor cambio es extraversión mínima, −1,71, y después neuroticismo máximo, +1,03. Los héroes quedan cerca del perfil asistente por defecto y los villanos se dispersan más. Esto demuestra sensibilidad de respuestas de autoinforme a descripciones que contienen el rasgo objetivo, no representación precisa de personas o poblaciones humanas: no hay datos humanos objetivo, validación conductual, criterio externo ni evaluación de fidelidad de personaje. La principal sobreextensión está en llamar «internal consistency reliability» a robustez agregada frente a perturbaciones. El propio artículo reconoce que sus transformaciones impiden calcular alfa de Cronbach y que la fiabilidad no implica validez. En consecuencia, el aporte sólido es un mapa amplio de sensibilidad del BFI a formatos, idiomas y personas instruidas; no valida el BFI como medida de personalidad de LLM ni respalda sustituir participantes humanos.

Research question

To what extent do BFI scores of various LLMs remain stable under changes in instruction, paraphrase, language, format, and order, and how much can GPT-3.5 shift those scores through context, profiles, or instructed personas?

Method

Full factorial design of 2,500 configurations per model on the BFI: five templates, five item versions, ten languages, five labels, and two orders. At temperature 0, OCEAN scores, PCA projections, DBSCAN outliers, mean differences, t-tests, and variance comparisons are computed. GPT-3.5 is additionally queried biweekly over approximately ten time points. For steerability, English variants are repeated with emotional narratives, extreme profiles, and personas, with or without an additional description called CoT.

Sample: Four LLMs evaluated across 2,500 BFI configurations each, at temperature 0. The temporal study of GPT-3.5 covers approximately ten biweekly measurements and two snapshots. The manipulation experiments are limited to GPT-3.5 and generate approximately 2,500 points per method according to the text. There are no new human participants or a human sample matched to the simulated persons.

Findings

  • In GPT-3.5, 77 of 2,500 configurations (3.08%) are DBSCAN outliers; they are associated mostly with numeric labels, descending order, and Arabic or Chinese.
  • Only 7/135 differences of one level versus the remaining exceed 0.15, although many comparisons are statistically significant and the 0.15 cutoff is not validated.
  • The OCEAN means of GPT-3.5 are 4.31, 4.15, 3.89, 4.13, and 2.35; the deviations 0.44, 0.39, 0.43, 0.38, and 0.42 are smaller than the compared external human norms.
  • GPT-4-Turbo, Gemini-1.0-Pro, and LLaMA-3.1-8B have 5.6%, 4.2%, and 4.4% outliers; LLaMA shows the most decentralized distribution.
  • Table 2 of the test-retest contradicts the global conclusion: the maximum of Agreeableness versus its mean has p shown as 0.00 and is flagged as unequal mean.
  • Emotional narratives barely shift BFI; only portray, among QA/BIO/POR, clearly alters the distribution, while the additional CoT-type description shows no visible effect.
  • Extreme profiles significantly change the target dimension; the largest differences from the default are -1.71 in minimum extraversion and +1.03 in maximum neuroticism.
  • Heroes resemble the default profile and villains are more dispersed, but there is no external reference for character fidelity.

Limitations

  • Paraphrases and translations modify an instrument validated for humans; only a sample of translations is reviewed and each cultural or linguistic version is not revalidated.
  • The study does not compute Cronbach's alpha, omega, item-total correlations, or any other measure of internal consistency; the authors acknowledge that the transformations preclude using alpha.
  • Concentration of scores at temperature 0 and lower variance than a heterogeneous human population do not distinguish stable personality from determinism, acquiescence, social desirability, or format bias.
  • The 2,500 configurations share templates, items, and transformations and are not independent participants; the massive t-tests do not correct for multiplicity or model the factorial dependence.
  • The analysis uses a descriptive criterion of a 0.15 difference without justifying its psychometric relevance, while ignoring numerous small effects with low p-values.
  • The test-retest selects minima and maxima post hoc, provides no stability coefficient or intervals, confuses failure to reject with accepting equality, and contains a significant exception that contradicts the text.
  • The human comparison uses published norms, not an equivalent sample subjected to the same translations, orders, formats, and conditions.
  • The manipulation evaluates only GPT-3.5 using the same self-report that the instruction attempts to change; it does not measure behavior, generalization to another test, or persistence beyond the prompt.
  • There is no human evaluation of profiles or personas, fidelity metrics, target demographic data, or contrast with real distributions of groups.

What the study does not establish

  • It does not validate the BFI as a measure of psychological personality, internal construct, or behavior of an LLM.
  • It does not demonstrate internal consistency of the instrument despite using that label in Findings 1.
  • It does not demonstrate that lower variance implies greater reliability or that non-random responses correspond to traits.
  • It does not demonstrate complete temporal stability: a comparison of Agreeableness contradicts the conclusion and only GPT-3.5 is studied over a few months.
  • It does not demonstrate precise understanding of profiles; repeating differences explicitly suggested by the prompt may be simple instruction following.
  • It does not demonstrate that GPT-3.5 accurately represents diverse human populations or that it can substitute participants in social sciences.
  • It does not establish convergent, discriminant, criterion, predictive, behavioral, or cross-cultural validity.

Traceability

Scope: Full text

Version: EMNLP 2024 main proceedings, ACL Anthology 2024.emnlp-main.354

Consulted source: https://aclanthology.org/2024.emnlp-main.354.pdf

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5-Turbo-1106
  • GPT-3.5-Turbo-0613 (test-retest period)
  • GPT-4-Turbo-1106
  • Gemini-1.0-Pro
  • LLaMA-3.1-8B
  • GPT-4-Turbo (item paraphrasing)

Instruments and metrics

  • 44-item Big Five Inventory (BFI)
  • Five instruction templates
  • Five item phrasings
  • Ten-language machine-translation set
  • Five choice-label formats and two orders
  • PCA visualization
  • DBSCAN outlier detection (eps 0.3, minPts 20)
  • Student t-tests and variance comparisons

Data used

  • 2,500 factorial BFI prompt configurations per evaluated model
  • Biweekly GPT-3.5 BFI measurements from September 2023 to January 2024
  • Sixteen emotional contexts, ten extreme trait profiles and sixteen named characters
  • External human BFI norms from Srivastava et al. (2003)

Evidence and location

  • Objectives, distinction between reliability and validity, and claim to represent groups: EMNLP 2024, pp. 6152-6154, Abstract, Introduction and section 2.2
  • Factorial of 2,500 configurations, translations, and item versions: EMNLP 2024, pp. 6154-6155, section 3.1 and Table 1
  • GPT-3.5 results, outliers, comparisons, and human norms: EMNLP 2024, pp. 6155-6156, section 3.2, Figure 1 and Table 9
  • Test-retest and Agreeableness contradiction: EMNLP 2024, p. 6157, section 3.3, Figure 2 and Table 2
  • Contexts, profiles, personas, and the effect of portray: EMNLP 2024, pp. 6157-6159, section 4 and Figures 3-4
  • Stated limitations on translation, alpha, and validity: EMNLP 2024, pp. 6159-6160, sections 5.1 and 6
  • Results for GPT-4, Gemini, and LLaMA: EMNLP 2024, pp. 6164-6165, Appendix A, Table 3 and Figures 5-7
  • Magnitudes of extreme profiles: EMNLP 2024, p. 6166, Appendix B, Figure 8 and Table 4
  • Prompts and detailed results of settings, profiles, and personas: EMNLP 2024, pp. 6167-6173, Appendices C-F and Tables 5-15