Personality Without Persons? A Psychometric Critique of Big Five Testing in Large Language Models

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Kim Zierahn, Cristina Cachero, Anna Korhonen, Nuria Oliver

Keywords: Personality, Psychometrics, Big Five, Construct validity, Measurement invariance

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

4
Authors
8
Findings
10
Limitations
6
Evidence

Editorial summary

English

This preprint, described by its repository as an AIES 2026 submission, asks whether Big Five questionnaires support personality attributions to LLMs that are comparable to human traits. It examines three requirements: whether items appropriately describe chatbots, whether scores differentiate models, and whether responses reproduce the expected five-factor latent structure. In phase one, three experts, two of them authors, rated five inventories. The LLM-adapted BFI-LLM and LMLPA passed the chatbot-suitability threshold, whereas BFI-44, IPIP-NEO-120, and MPI did not. A pilot with nine models, two finalist inventories, seven prompt templates, and three repetitions selected BFI-LLM and a template that explicitly frames the task as a psychological evaluation. In phase two, 44 items were sent separately to 244 model entries from 49 families, using two visual orders of the same 1–5 mapping and five repetitions per order. The released repository contains 107,352 deduplicated BFI rows, and its preprocessing reconstructs the reported 106,058 valid responses. Means cluster toward socially desirable endpoints: high Agreeableness, Conscientiousness, Extraversion, and Openness and low Neuroticism. A mixed model assigns 3% of variance to a general model intercept, 37% to item, 32% to model-by-item interaction, and 28% to residual variation. Robust CFA on the 244 × 44 matrix fits very poorly; Openness, Conscientiousness, Extraversion, and Agreeableness correlate at .92–.99, while reverse-keyed items load much more weakly than positive items. A two-factor EFA also fits inadequately. This is the strongest finding: under this self-report protocol, the human Big Five structure is not recovered and trait labels lack demonstrated structural validity. The conclusion must be narrower than some of the paper's prose. The 3% term represents each model's general response level and does not erase the 32% model-item interaction, so it does not prove that models lack stable differences. Human reference values come from the original BFI while model values use adapted stems, with no measurement-invariance or scale-linking study; they are directional context, not directly equivalent norms. Nineteen base/instruction pairs shift in a socially desirable direction on four traits, but the comparison mixes inference routes, providers, templates, and defaults; it suggests an association with instruction tuning rather than causally identifying alignment as the primary driver. The repository is unusually substantial, with raw responses and analysis code, but it is not one-command reproducible: preprocessing writes `final_df` while R reads `final_dfs`, derived files, an R environment, tests, CI, a license, and run instructions are absent. The parser also takes the first standalone 1–5 digit from many explanatory responses. A sensitivity rule prioritizing a later explicit final answer changes 2,369 BFI rows across 92 models. Global means move little, but an individual model-trait score changes by as much as 1.17 points. The aggregate pattern is therefore informative, while model rankings and family, country, or subgroup contrasts should not be treated as validated personality measurements.

Español

Este preprint y envío declarado a AIES 2026 pregunta si los cuestionarios Big Five permiten atribuir a los LLM diferencias de personalidad comparables a las humanas. El estudio cubre tres requisitos: que los ítems describan de forma apropiada a un chatbot, que las puntuaciones distingan modelos y que las respuestas reproduzcan la estructura latente de cinco factores. En la primera fase, tres expertos, dos de ellos autores, evaluaron cinco inventarios. Los dos adaptados a LLM, BFI-LLM y LMLPA, superaron el umbral de adecuación para chatbots; BFI-44, IPIP-NEO-120 y MPI no. Un piloto con nueve modelos, dos inventarios finalistas, siete plantillas y tres repeticiones seleccionó BFI-LLM y una plantilla que identifica la tarea como evaluación psicológica. En la segunda fase se administraron por separado 44 ítems a 244 entradas de modelos de 49 familias, con dos órdenes visuales de la misma escala 1–5 y cinco repeticiones por orden. De 107.352 filas BFI deduplicadas que contiene el repositorio, el preprocesado reproduce las 106.058 respuestas válidas declaradas. Las medias se concentran en los extremos socialmente deseables: alta amabilidad, responsabilidad, extraversión y apertura, y bajo neuroticismo. El modelo mixto atribuye 3% de la varianza a un intercepto general por modelo, 37% al ítem, 32% a la interacción modelo–ítem y 28% al residual. La CFA robusta sobre la matriz 244 × 44 presenta ajuste muy deficiente; apertura, responsabilidad, extraversión y amabilidad correlacionan entre .92 y .99, y los ítems invertidos cargan mucho peor que los positivos. Una EFA de dos factores tampoco ajusta bien. Este es el hallazgo más sólido: bajo este protocolo de autorreporte, no se recupera la estructura Big Five humana y las etiquetas de rasgo no tienen validez estructural demostrada. La conclusión debe ser más estrecha que parte del texto. El 3% corresponde al nivel general de respuesta por modelo y no elimina la interacción modelo–ítem del 32%; por tanto, no prueba que los modelos carezcan de diferencias estables. Las comparaciones con normas humanas usan ítems adaptados y no disponen de invariancia o equiparación de medida, de modo que son referencias direccionales, no una comparación normativa directa. Los 19 pares base/instruction muestran desplazamientos socialmente deseables en cuatro rasgos, pero mezclan rutas de inferencia, proveedores, plantillas y valores por defecto; sugieren una asociación con instruction tuning, no identifican causalmente el alineamiento como motor principal. La auditoría del repositorio encuentra datos y código sustanciales, pero no reproducción de un comando: el preprocesado escribe en `final_df` y R lee `final_dfs`, faltan derivados, entorno R, pruebas, CI, licencia e instrucciones de ejecución. Además, el parser toma el primer dígito 1–5 de muchas explicaciones. Una sensibilidad que prioriza una respuesta final explícita cambia 2.369 filas BFI de 92 modelos: las medias globales apenas varían, pero una puntuación individual puede cambiar hasta 1,17 puntos. Por ello, el patrón agregado es informativo, mientras que rankings por modelo, diferencias de familia, país o subgrupo no deben tratarse como medidas de personalidad validadas.

Research question

Do Big Five applications to LLMs meet three basic psychometric requirements: adequate content to describe chatbots, interpretable differences between models, and an internal structure coherent with the five human factors?

Method

Psychometric study in two phases. Three experts evaluated 252 items from five inventories using S-CVI and Gwet AC2; a factorial pilot with nine models compared BFI-LLM and LMLPA under seven templates. The main phase administered 44 BFI-LLM items separately to 244 model configurations, with two visual orders and five repetitions per order. It analyzed descriptives, an orientative comparison with human norms, a crossed mixed linear model, alpha, omega, robust CFA, parallel analysis/EFA, and exploratory comparisons by size, date, reasoning, license, origin, family, and 19 base/instruction pairs.

Sample: Content phase: three experts, two authors. Pilot: nine LLMs. Main phase: 244 model entries, 49 families; 112 from US companies, 98 Chinese, 21 French, four from UAE, and nine from other origins; 109 proprietary and 135 open-weight; 92 labeled as reasoning. The repository contains 107,352 deduplicated BFI rows and 106,058 responses that its parser considers valid.

Findings

  • BFI-LLM and LMLPA exceed the panel's adequacy threshold for chatbots; BFI-44, IPIP-NEO-120, and MPI do not, although the small and partially author panel limits the independence of this conclusion.
  • Model scores shift toward socially desirable extremes and show lower dispersion than published human references.
  • The reported decomposition is 3% model, 37% item, 32% model-item interaction, and 28% residual; the 3% measures general elevation, while the interaction preserves specific differences per item.
  • The five-factor CFA fits very poorly (CFI .53, TLI .50, RMSEA .17, SRMR .33) and four factors correlate .92-.99; a two-factor EFA also does not achieve adequate fit.
  • Reverse-worded items show much lower loadings than positive ones, compatible with a strong wording/negation effect and not with five separable traits.
  • Nineteen base/instruction pairs change in a socially desirable direction for four traits, but Neuroticism is not significant in the public paired t-test and the design does not identify a causal effect of alignment.
  • The independent reproduction confirms the count of 106,058 valid responses and the overall means, but a parser sensitivity modifies 2,369 BFI responses from 92 models and strongly alters some individual profiles.
  • The public artifact allows substantial audit, but the R analyses do not run from a clean clone without rebuilding derivatives and correcting the `final_df`/`final_dfs` path.

Limitations

  • Only content, distribution, and internal structure are tested under an English template; there is no temporal stability, cross-situational consistency, or prediction of real behavior.
  • The BFI-44 human norms and the adapted BFI-LLM items are not equated or subjected to measurement invariance, so their means and dispersions are not directly comparable.
  • The 244 records are not independent units: they include checkpoints, variants, and families that share architecture and data; the CFA and most comparisons do not model this dependence.
  • The first-digit parser can confuse numbers mentioned in explanations with the final response and especially affects results by model and subgroup.
  • The code normally averages ten responses per item, five per order, while the text describes an aggregation of five repetitions.
  • The KS test uses normalized parameters estimated on the same sample without Lilliefors correction; its normality p-values are not formal calibrated inference.
  • The base/instruction comparison mixes providers, chat templates, JSON constraints, and default values, and some pairs change in the opposite direction to the average.
  • The geographic and family analyses are confounded with company, architecture, size, license, provider, and alignment; France corresponds solely to Mistral.
  • The repository lacks a fixed R environment, tests, CI, license, reproduction guide, and required derivatives; it uses API aliases and mutable provider defaults.
  • The text confuses Q1 with Q3 for Extraversion and Openness and labels responses as scores aggregated by model.

What the study does not establish

  • It does not demonstrate that LLMs lack any stable behavioral dimension or differences between models; it demonstrates that this protocol does not validate the human Big Five structure.
  • It does not demonstrate that the 32% model-item interaction is irrelevant or that the 3% general intercept summarizes all differentiation between models.
  • It does not establish quantitative equivalence between chatbot scores and human norms or allow valid human percentiles for models.
  • It does not causally identify instruction tuning or alignment as the main driver of socially desirable scores.
  • It does not validate personality rankings of models, families, companies, countries, or regions.
  • It does not measure culture or national personality; the country is a company-origin label strongly confounded.
  • It does not prove that the adapted items have stability, criterion validity, or relevance for behavior outside the self-report task.
  • It does not yet constitute an executable reproduction from a command despite publishing extensive data and scripts.

Traceability

Scope: Full text

Version: arXiv:2607.02325v1; AIES 2026 submission, not an identified acceptance

Consulted source: https://arxiv.org/pdf/2607.02325

Review: Codex 19-page full-text visual, TeX, publication, psychometric, raw-data, parser, code and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • 244 model entries from 49 families
  • 19 matched base and instruction-tuned pairs
  • Nine-model pilot sample

Instruments and metrics

  • BFI-44
  • IPIP-NEO-120
  • Machine Personality Inventory (MPI)
  • BFI-LLM
  • Language Model Linguistic Personality Assessment (LMLPA)
  • 15-item social-desirability inventory collected in the repository but not central to the manuscript

Data used

  • Official BigFive-LLM-Evaluation repository: 244 large-scale response CSVs
  • Nine pilot-response CSVs
  • Expert item-rating dataset
  • Model metadata for 244 entries
  • McConochie 2007 human BFI-44 normative reference

Evidence and location

  • Preprint status, authors, date, and declared scope: Official arXiv record 2607.02325v1 and 19-page PDF, checked 2026-07-16
  • Content, pilot, and main design: arXiv v1, Sections 4-5 and Appendices A-C
  • Variance, reliability, CFA, EFA, and base/instruction pairs: arXiv v1, Results and Appendices F-G
  • Recognized limitations: arXiv v1, Discussion and Limitations
  • Raw data, parser, paths, notebooks, and counts: Official repository ellisalicante/BigFive-LLM-Evaluation at commit 62ab01bdc7335c372593a011961501e4d58eb438
  • Independent parser sensitivity and consolidated audit: reports/verification/article-283-aies-submission-big-five-content-validity-parser-variance-factor-reproducibility-and-claim-audit.json