Challenging the Validity of Personality Tests for Large Language Models

Evaluation and psychometric validity2025research-collection.ethz.chApproved editorial review

Authors: Tom Sühr, Florian E. Dorner, Samira Samadi, Augustin Kelava

Keywords: Computation and Language, Artificial Intelligence, Machine Learning, Personality assessment

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

4
Authors
9
Findings
14
Limitations
12
Evidence

Editorial summary

English

This paper raises a foundational psychometric objection to administering human questionnaires to language models and interpreting their scores as equivalent traits. It evaluates GPT-3.5, GPT-4, Llama 2 70B Chat, and Llama 3.1 70B Instruct in two studies. The first compares acquiescence bias on the 50-item IPIP Big Five Markers with a distribution of more than one million human responses. The second generates BFI-2 questionnaires under persona, cumulative-context, and seeded-first-answer conditions, explores their components with varimax-rotated PCA, and tests factor models with CFA. Its most important contribution is the separation of internal consistency from validity: a high Cronbach's alpha or omega cannot support interpretation when the structural model that gives those coefficients meaning fits poorly.

All four models agree with both positively and negatively keyed items more than the human mean, at approximately 0.6 for GPT-4, 0.8 for Llama 3.1, and 1.5 for GPT-3.5 and Llama 2. Only the latter two lie outside the empirical human distribution at p < .005; GPT-4 and Llama 3.1 are not statistically significant, although GPT-4 exceeds 89% of human agree-bias values. The comparison treats each model as one response profile rather than a sample of people or model runs, so it measures how unusual that output is relative to humans without estimating between-run model uncertainty.

For the BFI-2, Llama 3.1 under cumulative context and persona prompts produces the most human-looking exploratory pattern: five clearer blocks with opposite signs for direct and reverse-keyed items. That appearance is not confirmed. Every model and prompt setting remains far from the stated CFI/TLI ≥ .95 and RMSEA ≤ .06 thresholds. In the BFI-2 model, for example, GPT-4 reaches CFI .88/.78, TLI .81/.65, and RMSEA .15/.17; Llama 3.1, despite alpha and omega of .98/.93, reaches CFI .74/.66, TLI .65/.57, and RMSEA .38/.41. The solutions converge with lavaan warnings, including negative or zero estimated variances. The evidence strongly rejects a human-equivalent interpretation of BFI-2 scores for these models and protocols, although it is not a full formal multigroup invariance sequence: failure occurs at the prior requirement that the human configural model fit each response set adequately.

The induced population also constrains the inference. The 100 PersonaChat descriptions were not designed to span the Big Five space in a balanced way, and the recovered covariance structure depends on the range and correlations of traits induced by those prompts. Poor fit may reflect invalid instrument transfer, persona distribution, deterministic response mechanics, or a combination; it does not prove the absence of every form of personality or behavioral regularity in an LLM. Repository auditing finds only partial, manual reproducibility: there is no README, license, locked environment, or raw human dataset; the R scripts hard-code a local working directory and repeatedly overwrite their input objects; and the artifact predates the EAAMO version and contains no Llama 3.1 results. OpenAI outputs are also converted to one-hot answers from a returned token, whereas Llama uses an expected score over normalized probabilities for five option tokens, an asymmetry that can affect variance and estimated factor structure. The defensible conclusion is not that LLMs “have no personality,” but that these BFI-2 scores should not be treated as human traits until the construct, administration protocol, and relationship to external behavior have been validated.

Español

Este trabajo formula una objeción psicométrica fundamental a la práctica de administrar cuestionarios humanos a modelos de lenguaje y leer sus puntuaciones como rasgos equivalentes. Evalúa GPT-3.5, GPT-4, Llama 2 70B Chat y Llama 3.1 70B Instruct en dos estudios. El primero compara el sesgo de aquiescencia en los 50 ítems IPIP Big Five Markers con una distribución de más de un millón de respuestas humanas. El segundo genera cuestionarios BFI-2 bajo personas, contexto acumulado y respuestas iniciales sembradas, explora sus componentes mediante PCA con rotación varimax y contrasta modelos factoriales con CFA. Su aportación más importante es separar consistencia interna de validez: un alfa de Cronbach o un omega alto no permite interpretar una escala si el modelo estructural que da sentido a esos coeficientes ajusta mal.

Los cuatro modelos muestran más tendencia a asentir a ítems de clave positiva y negativa que la media humana, con valores aproximados de 0,6 para GPT-4, 0,8 para Llama 3.1 y 1,5 para GPT-3.5 y Llama 2. Solo estos dos últimos se sitúan fuera de la distribución humana con p < 0,005; GPT-4 y Llama 3.1 no alcanzan significación, aunque GPT-4 supera al 89 % de la muestra humana. El contraste trata cada modelo como un único perfil de respuesta, no como una muestra de personas o ejecuciones, por lo que cuantifica lo inusual de esa salida respecto de humanos sin estimar incertidumbre entre ejecuciones del modelo.

En el BFI-2, Llama 3.1 bajo contexto acumulado y personas produce el patrón exploratorio más parecido a cinco bloques, con signos opuestos para ítems directos e inversos. Esa apariencia no se confirma: todos los modelos y configuraciones quedan lejos de los umbrales declarados de CFI/TLI ≥ 0,95 y RMSEA ≤ 0,06. En el modelo BFI-2, por ejemplo, GPT-4 alcanza CFI 0,88/0,78, TLI 0,81/0,65 y RMSEA 0,15/0,17; Llama 3.1, pese a alfa y omega de 0,98/0,93, obtiene CFI 0,74/0,66, TLI 0,65/0,57 y RMSEA 0,38/0,41. Las soluciones convergen con advertencias de lavaan, incluidas varianzas negativas o nulas. La evidencia es fuerte para rechazar una interpretación humana del BFI-2 en estos modelos y protocolos, aunque no constituye una prueba formal completa de invariancia multigrupo: el fallo ocurre ya en la condición previa de que el modelo configural humano ajuste adecuadamente a cada conjunto de respuestas.

La población artificial también limita la inferencia. Las 100 descripciones de PersonaChat no fueron diseñadas para cubrir de forma equilibrada el espacio Big Five, y la estructura factorial observada depende del rango y las correlaciones de los rasgos inducidos por esos prompts. Un mal ajuste puede reflejar la transferencia inadecuada del instrumento, la distribución de personas, el mecanismo determinista de respuesta o una combinación de ellos; no prueba la ausencia de toda forma de personalidad o regularidad conductual en un LLM. La auditoría del repositorio encuentra una reproducción parcial y manual: faltan README, licencia, entorno bloqueado y datos humanos crudos; los scripts R fijan una ruta local y sobrescriben repetidamente entradas; y el artefacto, publicado antes de la versión EAAMO, no contiene Llama 3.1. Además, OpenAI se convierte a respuestas one-hot a partir de un token, mientras Llama usa el valor esperado de probabilidades normalizadas sobre cinco tokens, una asimetría que puede afectar la varianza y la estructura estimada. La conclusión defendible no es que los LLM «no tengan personalidad», sino que estas puntuaciones BFI-2 no deben tratarse como rasgos humanos sin validar primero el constructo, el protocolo y su relación con conducta externa.

Research question

Do the IPIP-50 and BFI-2 retain their psychometric properties and interpretation when applied to LLMs, or do acquiescence bias and a different factor structure prevent equating their scores with human traits?

Method

Study 1 administers the 50 IPIP Big Five Markers once without persona or repetition to four models and calculates, after reverse-scoring the negatively keyed items, the difference between the mean of positive and negative items. Each value is placed in the empirical distribution of 1,015,342 human participants. Study 2 administers the 60 BFI-2 items and collects 100 questionnaires per model and configuration: PersonaChat personas with each item in a new context, the same personas with the 60 items accumulated in one context, or accumulated context without persona seeding the first response between 1 and 5. The conditions without persona or seed lack sufficient variance. Standardized PCA is performed with five components and varimax rotation; then, CFA per domain for one component, three facets, and the hierarchical BFI-2 model, reporting alpha, hierarchical omega, CFI, TLI, and RMSEA. The review checked the complete preprint, the EAAMO 2025 publication, appendices, tables, figures, and the public repository.

Sample: The IPIP analysis compares a single profile of 50 responses per LLM with 1,015,342 human questionnaires, not multiple runs of each model. For BFI-2, n = 100 complete questionnaires are collected per model and analyzable configuration. The configurations with personas reuse 100 PersonaChat descriptions; the seeded configuration fixes the first response at one of five values and continues the 59 items in the same context. The conditions without persona or seed do not achieve sufficient variance even when increasing temperature. The publication includes four models, but the March 2024 repository only contains artifacts for GPT-3.5, GPT-4, and Llama 2.

Findings

  • Acquiescence bias is approximately 0.6 for GPT-4, 0.8 for Llama 3.1, and 1.5 for GPT-3.5 and Llama 2; only GPT-3.5 and Llama 2 differ from the human distribution with p < 0.005.
  • GPT-4 exceeds the acquiescence bias of 89% of humans, but neither GPT-4 nor Llama 3.1 reaches statistical significance in the published test.
  • Without persona or seeded initial response, most items barely vary, even at higher temperatures, preventing useful PCA and CFA estimation.
  • In context with personas, Llama 3.1 shows the clearest exploratory separation in PCA between five groups and between direct and reverse items; GPT-4 shows limited separation and GPT-3.5/Llama 2 do not reproduce it.
  • The promising structure of Llama 3.1 does not pass CFA. None of the four models simultaneously achieves CFI/TLI >= 0.95 and RMSEA <= 0.06 under the BFI-2 model.
  • All mean alpha and omega values are above 0.70 (Llama 3.1 reaches 0.98/0.93), but their interpretation as reliability is not valid when the underlying factor model fits poorly.
  • The CFA solutions converge with lavaan warnings, including negative or zero variances, a sign of improper or unstable solutions that requires interpreting parameters and indices with caution.
  • The most robust result is a burden of proof: those who apply a human scale to LLMs must demonstrate its validity for that construct, model, and protocol, rather than assuming it from plausible scores.
  • The two experiments employ distinct instruments (IPIP-50 for acquiescence and BFI-2 for structure), so their results are complementary, not a single integrated test of invariance.

Limitations

  • The acquiescence test uses a single deterministic output per model; the reference distribution is human and does not quantify variability across seeds, sessions, dates, or LLM runs.
  • Significance is obtained only for GPT-3.5 and Llama 2. Describing all four models as statistically non-human would unduly extend the result.
  • The 100 PersonaChat personas are not sampled or balanced to produce a known Big Five distribution. Factor identification depends on the range, covariance, and representativeness of those manipulations.
  • The study does not run a formal sequence of configural, metric, scalar, and residual multigroup invariance; it demonstrates first that the human model does not even offer sufficient configural fit in the LLM datasets.
  • PCA with varimax imposes orthogonal components, whereas Big Five domains need not be perfectly independent; other rotations or models could change the exploratory representation.
  • The table averages CFA indices over five domains or models per facet and does not expose uncertainty or all individual results in the main body, which may hide heterogeneity across traits.
  • Negative or zero variances indicate improper solutions. Although they are consistent with instability and poor fit, they also limit the precision with which the estimated parameters themselves can be interpreted.
  • Convergent or predictive validity against external behavior, test-retest stability, changes across versions, real tasks, languages, or robustness to paraphrases and formats are not evaluated.
  • The models are snapshots from 2023 along with Llama 2 and Llama 3.1; the result does not automatically transfer to later systems or other response modalities.
  • Using an LLM as an individual in the acquiescence analysis and as a population conditioned by personas in the factor analysis are different methodological choices and do not resolve what ontological unit a synthetic personality should measure.
  • The article reuses the OpenPsychometrics IPIP sample, but the repository does not include raw human responses or a complete and documented cleaning and quality-control pipeline.
  • The repository lacks a README, license, and fixed dependencies; the R scripts contain an absolute local path, overwrite objects from different sources, and require manually activating blocks and paths.
  • The public code precedes the EAAMO version and does not reproduce Llama 3.1. It also uses different representations per provider: a one-hot response for OpenAI and a probabilistic mean for Llama, potentially relevant for factor variance.
  • The preprint declares zero temperature for OpenAI, while several public calls use 1.0; given that the condensed proceedings version does not detail all parameters, this is documented as a discrepancy between preprint and artifact, not as an unequivocal contradiction of the final method.

What the study does not establish

  • It does not demonstrate that LLMs lack any personality, disposition, style, or useful behavioral regularity; it invalidates a specific human interpretation of the BFI-2 under the tested protocols.
  • It does not demonstrate that all psychological questionnaires always fail on all models; each instrument, construct, model, and protocol needs its own evidence.
  • It does not demonstrate the absence of five dimensions in all model behavior, but rather the lack of fit of the human factor model in the responses induced and observed here.
  • It does not allow attributing acquiescence bias to RLHF, training, architecture, or annotator preferences; those explanations are proposed but not experimentally manipulated.
  • It does not turn a questionnaire score into evidence of subjective experience, identity, intention, self-knowledge, or internal psychological states.
  • It does not establish that alpha or omega are useless in general; it shows that they cannot be interpreted as reliability when the structural assumptions they require fail.
  • It does not yet offer a validated alternative instrument for LLM-specific personality; it presents criteria and a construction program that requires external validation.
  • It does not prove that Llama 3.1 possesses a Big Five structure because its PCA is visually more orderly; the subsequent CFA rejects sufficient fit to the human model.

Traceability

Scope: Full text

Version: arXiv:2311.05297v2 (5 Jun 2024); EAAMO 2025 proceedings, DOI 10.1145/3757887.3763016, also reviewed (published PDF SHA-256 d291dc59941901d1a99671a9a586a06595116cddbf690c70d371b28ec26e147f); public code commit 7ab6bfccd183ab1bf19857064d84388a282338d2 audited

Consulted source: https://arxiv.org/pdf/2311.05297

Review: Codex editorial review and methods/code audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI gpt-3.5-turbo-0613
  • OpenAI gpt-4-0613
  • Meta Llama-2-70b-chat
  • Meta Llama-3.1-70B-Instruct

Instruments and metrics

  • 50-item IPIP Big Five Markers
  • Big Five Inventory-2 (BFI-2, 60 items and 15 facets)
  • Agree-bias statistic from true-key versus false-key items
  • Principal component analysis with varimax rotation
  • Confirmatory factor analysis with lavaan
  • Cronbach's alpha
  • McDonald's hierarchical omega
  • CFI, TLI and RMSEA model-fit indices

Data used

  • OpenPsychometrics IPIP Big Five human responses (n = 1,015,342)
  • PersonaChat subset of 100 persona descriptions
  • BFI-2 human component loadings from Soto and John (2017)
  • Online and student human BFI-2 samples reported by Soto and John (2017)
  • Released LLM questionnaire responses and preprocessed PCA/CFA tables for GPT-3.5, GPT-4 and Llama 2

Evidence and location

  • Definitive publication, authorship, DOI, license, and pagination: EAAMO 2025 proceedings paper, pp. 74-81; DOI 10.1145/3757887.3763016; ETH Research Collection 10.3929/ethz-c-000789617
  • Definition and results of acquiescence bias: EAAMO 2025 paper, section 4.3 and Figure 1, pp. 76-77
  • Models, BFI-2 configurations, n = 100, and absence of variance without persona or seed: EAAMO 2025 paper, section 4.4, p. 77; Appendix D in arXiv:2311.05297v2
  • PCA with varimax and comparison of loadings with human samples: EAAMO 2025 paper, section 4.4 and Figure 2, pp. 77-78
  • Interpretive dependence of alpha and omega on the factor model: EAAMO 2025 paper, section 4.5, pp. 78-79
  • Values of alpha, omega, CFI, TLI, and RMSEA for four models: EAAMO 2025 paper, Table 1, p. 80
  • lavaan warnings and conclusion on fit and invariance: EAAMO 2025 paper, section 4.6, p. 79
  • Scope, limitations, and proposal of LLM-specific instruments: EAAMO 2025 paper, sections 5-7, pp. 79-80
  • Sample sizes and provenance, prompts, and parameters from the preprint: arXiv:2311.05297v2, Appendix A-D and Tables 1-2
  • Partial reproducibility, missing data, local paths, and manual flow: GitHub tsuehr/LLMpersonality commit 7ab6bfccd183ab1bf19857064d84388a282338d2; analysis/cfa.R, analysis/pca.R and analysis/final_analysis_and_plots.ipynb
  • Temperature differences and response representation across providers: GitHub commit 7ab6bfccd183ab1bf19857064d84388a282338d2; experiment_general.py, experiment_incontext.py, experiment_incontext_seeded.py and Utils.py
  • Absence of Llama 3.1 in the public artifact prior to publication: GitHub commit 7ab6bfccd183ab1bf19857064d84388a282338d2 dated 5 Mar 2024; repository file tree and released result tables