A Comparative Study of Large Language Models and Human Personality Traits

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Wang Jiaqi, Wang bo, Guo fa, Cheng cheng, Yang li

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

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

5
Authors
23
Findings
48
Limitations
12
Evidence

Editorial summary

English

This 142-page manuscript compares responses from 60 young adults with repeated runs of ChatGLM3-6B, an ambiguously named DeepSeek-V3/R1 model, GPT-4o, and Llama3.1-8B on Chinese versions of BFI-2 and MBTI and author-created paraphrases. It presents three studies: test-retest stability, consistency across wording variants, and retention of an assumed baseline personality during role-play. It then proposes an Ecological Distribution-based Personality Definition and a Distributed Personality Framework, describing LLM scores by their mean and variance as dynamic, context-dependent outputs. The only located source is arXiv v1; no later publication, code, data, or complete materials were found. Study 1 does not perform a statistically commensurate comparison between humans and models. The 60 participants, all aged 18–30 with at least a bachelor's degree, were preselected to yield 30 MBTI E and 30 I participants. They answered seven forms, a 60-item BFI-2 plus three variants and a 93-item MBTI plus two variants, 519 responses per session, in one day and again two weeks later. Their dimension-level test-retest correlations range from 0.716 to 0.931. Each LLM instead completes the seven forms 100 times within one experimental setting and is summarized by means and variances, not test-retest correlations or a time interval. The manuscript directly compares these incompatible quantities and claims significant differences without a human-versus-LLM test. It also does not report the human individual-score distribution or variance that the prose says is nearly zero. Reported LLM variances range from approximately 0.97 to 14.62, with ChatGLM refusals and an invalid C response from Llama on A/B items. These observations do show sensitivity to model, format, and invalid-response handling. They do not by themselves establish instability of a psychological entity. The protocol simultaneously calls runs independent and says prompts and random seeds are strictly fixed; temperature, top-p, provider, date, snapshot, seed, endpoint, chat template, parser, and retry policy are absent. To support normality, the manuscript plots Excel NORM.DIST curves from means and standard deviations and then concludes that the results are normally distributed. The normal curve was imposed rather than tested against the observations. Study 2 uses the same seven forms and data to examine wording changes. Humans have reported single-measure ICCs of 0.881–0.938 and average-measure ICCs of 0.967–0.981; LLMs are again compared through means and variances. The variants are not independently validated instruments: they transform items into social comparison, external judgment, or sentence completion and can change the response scale. A high ICC may reflect stable between-person differences, while average-measure ICC mechanically rises when forms are averaged; neither establishes semantic, factorial, or psychometric equivalence. Again, there is no common human-versus-LLM consistency test, despite claims of significant differences and inferences about internal mechanisms from surface scores. Study 3 reuses the same 60 participants, who spend four days role-playing Lin Daiyu, Sun Wukong, a very introverted person, and a very extroverted person. Analyses split people by baseline MBTI E/I and apply independent-samples t tests to scores and deviations even though the design is declared within-subject and baseline and role-play measures are linked. Results are mixed by scale, variant, and role. For LLMs, four models, four roles, and ten tests per role are announced, yet the only published tables compare GPT with DeepSeek for Lin Daiyu, Sun Wukong, and very introverted; ChatGLM, Llama, and very extroverted are absent. The reported GPT-versus-DeepSeek effect sizes (d=0.921, 2.508, and 0.876) show differences between those outputs, not that GPT role-plays better or that model parameter count causes personality retention. Model identity prevents reproduction and attribution. DeepSeek-V3/R1 combines two distinct models and later becomes simply Deepseek, without identifying the executed model. GPT-4o has no snapshot. The text does not say whether ChatGLM3-6B and Llama3.1-8B are base or instruction-tuned. Chapter 5 also calls ChatGLM3-6B a 600-million-parameter model and Llama3.1-8B an 800-million-parameter model; official sources describe 6B and 8B models. The same chapter labels DeepSeek small, although Chapter 3 and official V3/R1 sources describe 671B total parameters. Model, scale, language, architecture, and access are fully confounded in a sample of four, so effects of size, culture, parameter count, or architecture cannot be isolated. The usable contribution is exploratory: under seven Chinese forms, unvalidated paraphrases, and an incomplete protocol, response distributions differ across models, wording, and roles, and refusals or invalid formats occur. It is reasonable to recommend reporting response distributions, prompt sensitivity, and error handling rather than a single personality score. The published evidence does not establish that LLMs possess psychological traits, that human scales are generally incompatible, that outputs are normally distributed, that larger models are more stable, that baseline personality causes role-play behavior, or that the distributed framework has been validated.

Español

Este manuscrito de 142 páginas compara respuestas de 60 adultos jóvenes con ejecuciones repetidas de ChatGLM3-6B, un modelo denominado de forma ambigua “DeepSeek-V3/R1”, GPT-4o y Llama3.1-8B en versiones chinas del BFI-2 y MBTI y en paráfrasis creadas por los autores. Presenta tres estudios: estabilidad test-retest, consistencia entre variantes de redacción y retención de una supuesta personalidad basal durante role-play. A partir de ellos propone una Ecological Distribution-based Personality Definition y un Distributed Personality Framework: describir las puntuaciones del LLM por su media y varianza, como salidas dinámicas dependientes del contexto. La única fuente localizada es arXiv v1, sin publicación posterior, código, datos ni materiales completos. El Estudio 1 no realiza una comparación estadística homogénea entre humanos y modelos. Los 60 participantes, todos de 18–30 años y con grado universitario o superior, fueron preseleccionados para obtener 30 E y 30 I según MBTI. Contestaron siete formularios, BFI-2 de 60 ítems más tres variantes y MBTI de 93 ítems más dos variantes, 519 respuestas por sesión, en un día y de nuevo dos semanas después. Sus correlaciones test-retest por dimensión van de 0,716 a 0,931. Cada LLM, en cambio, completa los siete formularios 100 veces bajo una misma sesión experimental y se resume mediante media y varianza, no con correlaciones test-retest ni intervalo temporal. El manuscrito compara directamente esas magnitudes incompatibles y afirma diferencias significativas sin presentar una prueba humano-versus-LLM. Tampoco publica la distribución o varianza individual humana que, según el texto, sería casi nula. Las ejecuciones de LLM reportan varianzas entre aproximadamente 0,97 y 14,62, rechazos de ChatGLM y una opción C inválida de Llama en preguntas A/B. Eso sí muestra sensibilidad al modelo, al formato y al tratamiento de respuestas inválidas. No demuestra por sí solo inestabilidad de una entidad psicológica. El protocolo dice simultáneamente que cada ejecución es independiente y que prompt y semilla aleatoria quedan estrictamente fijos; no informa temperatura, top-p, proveedor, fecha, snapshot, seed, endpoint, plantilla de chat, parser o política de reintentos. Para justificar normalidad, el trabajo dibuja curvas NORM.DIST de Excel a partir de media y desviación estándar y luego concluye que los resultados siguen una distribución normal: la curva fue impuesta, no contrastada con los datos. El Estudio 2 usa los mismos siete formularios y datos para estudiar cambios de redacción. En humanos informa ICC de medida única de 0,881–0,938 e ICC de promedio de 0,967–0,981; en LLM vuelve a comparar medias y varianzas. Las variantes no son instrumentos validados independientes: convierten ítems en comparación social, juicio externo o completado de frases y modifican incluso la escala de respuesta. Un ICC alto puede reflejar diferencias estables entre personas y el ICC promedio crece mecánicamente al promediar formularios; no establece equivalencia semántica, factorial o psicométrica. De nuevo, no existe una prueba común de consistencia humana frente a LLM, aunque el texto habla de diferencias significativas e infiere mecanismos internos a partir de puntuaciones superficiales. El Estudio 3 reutiliza los mismos 60 participantes, que durante cuatro días representan a Lin Daiyu, Sun Wukong, una persona muy introvertida y una muy extrovertida. Los análisis separan a las personas por su MBTI E/I basal y aplican t de muestras independientes a puntuaciones y desviaciones, pese a que el diseño se declara intra-sujeto y las medidas basal/role-play están vinculadas. Hay resultados mixtos según escala, variante y rol. Para LLM se anuncian cuatro modelos y cuatro roles, diez tests por rol, pero las únicas tablas publicadas comparan GPT con DeepSeek para Lin Daiyu, Sun Wukong y muy introvertido; no aparecen ChatGLM, Llama ni muy extrovertido. Los tamaños de efecto reportados para GPT frente a DeepSeek (d=0,921; 2,508; 0,876) demuestran diferencias entre esas salidas, no que GPT tenga mejor role-play ni que el parámetro del modelo cause retención de personalidad. La identidad de los modelos impide reproducir y atribuir resultados. DeepSeek-V3/R1 agrupa dos modelos distintos y más adelante se reduce a Deepseek, sin aclarar cuál se ejecutó. GPT-4o no tiene snapshot. No se especifica si ChatGLM3-6B y Llama3.1-8B son base o instruct. Además, el capítulo 5 llama a ChatGLM3-6B un modelo de 600 millones y a Llama3.1-8B uno de 800 millones; las fuentes oficiales los describen como 6B y 8B. El mismo capítulo clasifica DeepSeek como pequeño aunque el capítulo 3 y las fuentes de V3/R1 lo describen como 671B total. Modelo, tamaño, idioma, arquitectura y acceso están completamente confundidos en una muestra de cuatro, por lo que no se puede aislar un efecto de escala, cultura, parámetros o arquitectura. La aportación utilizable es exploratoria: bajo siete formularios chinos, paráfrasis no validadas y un protocolo incompleto, las distribuciones de respuestas cambian entre modelos, redacciones y roles, y aparecen rechazos o formatos inválidos. Es razonable recomendar que las evaluaciones de personalidad de LLM reporten distribución, sensibilidad al prompt y manejo de errores en vez de una sola puntuación. Los datos publicados no establecen que los LLM posean rasgos psicológicos, que las escalas humanas sean incompatibles en general, que exista una distribución normal, que modelos grandes sean más estables, que la personalidad basal cause el role-play ni que el marco distribuido haya sido validado.

Research question

Do humans and various LLMs differ in the stability of scores, the consistency across paraphrases of BFI-2/MBTI, and the change in score when representing roles, and do those differences justify a specific distributed personality framework for LLMs?

Method

Three studies on Chinese versions of the BFI-2 and MBTI and five variants created by the authors. In humans, 60 participants complete seven forms twice with a two-week interval, then test-retest correlations are calculated; ICC is calculated between variants and, in another four-day protocol, E/I groups are compared across four roles using independent-samples t-tests. In LLMs, four model names complete each form 100 times with declared fixed prompt and seed; means and variances are calculated. For role-play, ten runs per role are announced, but only GPT versus DeepSeek is tabulated across three roles. No common statistic or direct human-versus-LLM test is presented.

Sample: The final human sample is 60 people recruited offline: 25 men, 35 women; 53 aged 18–24 and 7 aged 25–30; all with a university degree or higher. 30 classified as E and 30 as I are selected via MBTI, so it is not a representative natural sample. The text mentions pilots of 2 and about 10 participants, but does not clarify whether they are included in the 60, nor does it report exclusions, missing data, dropout, compensation, location, country, or approving institution. The model sample is nominally four, with 100 repetitions per scale in Studies 1–2 and ten per role in Study 3; the final role-play tables show only two models and three roles.

Findings

  • arXiv v1 of 1 May 2025 remains the only and latest local version found.
  • The document is a thesis-type manuscript of 142 pages, not a short peer-reviewed article.
  • The 142 pages were rendered and visually inspected; they contain no appendices of prompts, questionnaires, data, or code.
  • Humans complete 519 items per session and repeat the protocol two weeks later.
  • Human test-retest correlations per dimension range from 0.716 to 0.931, all labeled p<0.05.
  • LLMs complete the seven forms 100 times and are summarized via mean and variance.
  • Reported LLM variances span approximately 0.9724–14.6208.
  • ChatGLM rejects some items and Llama produces an invalid option C on A/B questions.
  • No direct statistical test comparing human stability with that of LLMs is published.
  • The human protocol uses correlations between two time points; the LLM protocol uses variance across repetitions with no interval.
  • Normality is illustrated with parametrically generated NORM.DIST curves, not with an empirical goodness-of-fit test or plot.
  • In humans, single-measure ICCs across variants range from 0.881–0.938 and average ICCs from 0.967–0.981.
  • The variants modify perspective, semantics, and in one case the response format, but receive no external validation.
  • The role-play study uses four roles and the same 60 humans across four days.
  • Human role-play results are mixed: some BFI contrasts are not significant and several MBTI ones are.
  • For LLMs, four models and four roles are announced, but only GPT and DeepSeek are tabulated across three roles.
  • Reported d values for GPT versus DeepSeek are 0.921 for Lin Daiyu, 2.508 for Sun Wukong, and 0.876 for highly introverted.
  • The manuscript interprets these d values as better GPT performance without defining or validating a role quality criterion.
  • DeepSeek-V3 and DeepSeek-R1 are grouped under a single experimental label and then it is not identified which was run.
  • Chapter 5 reports 600M for ChatGLM3-6B and 800M for Llama3.1-8B; official sources indicate 6B and 8B.
  • Chapter 5 calls DeepSeek small even though the official V3/R1 families have 671B total parameters.
  • No code, data, complete questionnaires, outputs, executable configuration, or associated repository were found.
  • The strongest contribution is documenting sensitivity to wording, model, role, and treatment of invalid responses.

Limitations

  • The main comparison pits human correlations against LLM variances, magnitudes that share neither scale nor interpretation.
  • There is no human-versus-LLM test, confidence interval for the difference, bootstrap, or hierarchical model.
  • The text claims significant differences where it only presents separate descriptive statistics.
  • Human stability is measured after two weeks; LLM repetitions have no comparable temporal dimension.
  • The 100 runs do not equal 100 independent subjects or a test-retest of the same psychological entity.
  • Prompt and seed are declared strictly fixed, but it is not explained what randomness produces the distributions or how the runs are independent.
  • Temperature, top-p, top-k, seed, provider, endpoint, snapshot, date, or software version are not reported.
  • Chat templates, system messages, item order, accumulated context, or conversation reset are not specified.
  • The parser, scoring, item reversal, retries, rejections, or imputation of invalid responses are not described.
  • ChatGLM rejections and Llama's option C may measure instruction/parsing failures rather than personality variation.
  • Raw or per-run results are not published, so distributions, outliers, or errors cannot be audited.
  • The NORM.DIST curve imposes normality from mean and standard deviation; it does not prove the data are normal.
  • Mean and variance are insufficient to justify a new psychological theory or an ecological distribution.
  • It is not operationally defined what observation would falsify the Distributed Personality Framework.
  • Human sampling is conditioned on 30 E and 30 I per MBTI, artificially altering between-person variance.
  • The human sample is small, young, highly educated, and without a clearly reported country or population frame.
  • It is not reported how participants were recruited, compensated, or retained, nor how many data were excluded.
  • Completing 519 items in one day and repeating them can introduce fatigue, learning, order, and memory.
  • Counterbalancing, order randomization, duration, attention, or response quality are not reported.
  • The variants were created by the authors and no back-translation, expert review, or sufficient cognitive piloting is provided.
  • The variants change perspective and response format, so they are not necessarily parallel forms.
  • There is no factor structure, measurement invariance, internal consistency, standard error, or convergent validity of the variants.
  • A high ICC may be dominated by between-subject variance and does not imply equivalence of means or content.
  • The average-measure ICC increases by construction when forms are added and should not be interpreted as independent validity.
  • BFI-2 and MBTI are combined as if they measured the same personality construct and with equal psychometric status.
  • Role-play is declared intra-subject, but independent-samples tests are used for linked measures.
  • Repeated measures, order of the four days, carryover, learning, or within-person dependence are not modeled.
  • Differences from a baseline are analyzed without propagating the measurement error of both scores.
  • Numerous t-tests are performed without correction for multiplicity or a clearly bounded primary hypothesis.
  • H0/H1 interpretations confuse non-significance with evidence of personality retention.
  • Lin Daiyu and Sun Wukong are not calibrated or validated points on a continuous extraversion scale.
  • There is no manipulation check or external evaluation of whether humans or LLMs portrayed the roles well.
  • No role-play gold standard is defined, so a score difference does not equate to better performance.
  • The LLM role-play study omits two of the four models and one of the four announced roles.
  • Ten repetitions per role are few for characterizing distributions and no uncertainty per seed is reported.
  • The names and numbers of tables 5-16/5-17 are inconsistent.
  • It is not identified which exact variant of DeepSeek was run; V3 and R1 are different models.
  • GPT-4o lacks a snapshot and ChatGLM/Llama lack an exact base/instruct identifier.
  • ChatGLM and Llama parameters are misstated by a factor of ten in chapter 5.
  • DeepSeek changes from large V3/R1 model in chapter 3 to small model in chapter 5.
  • Size, language, architecture, provider, and instruction-following capability are confounded across four models.
  • Four models do not allow inferring a general scaling trend or separating family size from training.
  • There is no preregistration, power analysis specific to the multiple contrasts, or published exclusion plan.
  • Ethics is limited to consent, anonymization, and declared support; no committee, approval, or protocol number is identified.
  • The limitations section does not acknowledge the statistical incompatibilities, result omissions, or central model errors.
  • There is no external replication, validation outside Chinese, new scales, behavioral tasks, or natural interaction.
  • It is not evaluated whether scores predict behavior, preferences, safety, dialogue quality, or human outcomes.
  • No peer review, code, data, complete prompts, or executable artifacts were located.

What the study does not establish

  • It does not establish that an LLM possesses a human personality or psychological structure.
  • It does not demonstrate a valid statistical difference between human stability and LLM stability.
  • It does not demonstrate that LLM outputs follow a normal distribution.
  • It does not demonstrate that mean and variance constitute a validated personality theory.
  • It does not demonstrate that larger models are more stable.
  • It does not isolate effects of language, culture, architecture, size, or training.
  • It does not demonstrate that the created variants are equivalent or valid instruments.
  • It does not demonstrate that LLMs understand the central meaning of items less well.
  • It does not demonstrate causality between a baseline personality of the model and its role-play.
  • It does not demonstrate that GPT portrays roles better than DeepSeek.
  • It does not generalize to other models, languages, populations, scales, or natural interactions.
  • It does not offer a reproducible result without checkpoints, configuration, data, and code.

Traceability

Scope: Full text

Version: arXiv:2505.14845v1, submitted 1 May 2025; 142-page English thesis-style manuscript; latest version confirmed through the arXiv API on 15 July 2026; no later proceedings or journal version located

Consulted source: https://arxiv.org/pdf/2505.14845v1

Review: Codex complete bilingual full-text fidelity pass, all-page PDF visual inspection, arXiv latest-version reconciliation, three-study design reconstruction, statistical comparability audit, model-identity and official-parameter cross-check, psychometric-validity assessment, result-omission and reproducibility review; summaries written from the complete 142-page manuscript rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • ChatGLM3-6B; exact checkpoint, base/instruct status and runtime not reported
  • DeepSeek-V3/R1, an unresolved conflation of distinct V3 and R1 models; later tables say only Deepseek
  • GPT-4o; provider is implied but model snapshot and execution date are not reported
  • Llama3.1-8B; exact checkpoint and base/instruct status not reported

Instruments and metrics

  • Chinese BFI-2, 60 items and five dimensions
  • BFI-2 Variant 1: social-comparison reformulation
  • BFI-2 Variant 2: external-description reformulation
  • BFI-2 Variant 3: sentence-completion/frequency reformulation
  • Chinese MBTI, 93 forced-choice items and four dichotomies
  • MBTI Variants 1 and 2: author-created reformulations of the two options
  • Pearson dimension-level test-retest correlation for humans
  • ICC(2,k) described for cross-form consistency, with single- and average-measure outputs
  • Mean and variance of 100 LLM runs
  • Excel NORM.DIST curves constructed from reported means and standard deviations
  • Independent-samples t tests, Welch tests, 95% confidence intervals and Cohen's d in role-play analyses

Data used

  • No public dataset or participant-level data located
  • No released LLM raw responses, score tables, invalid-response log, seeds or run identifiers
  • No complete released inventory variants, role schedule, prompts, scoring code or analysis code

Evidence and location

  • Version, date, authorship, and nature of manuscript: arXiv API record 2505.14845v1 checked 15 July 2026; manuscript cover, abstract and contents, pages I–IX
  • Human sample, scales, and test-retest protocol: Manuscript Sections 3.4.1–3.4.6, PDF pages 49–52; Tables 3-1 to 3-3
  • Models, 100 runs, fixed prompt and seed: Manuscript Sections 3.5.1–3.5.4, PDF pages 52–56; Table 3-4
  • Means, variances, rejections, and invalid option: Manuscript Tables 3-4 to 3-11 and discussion, PDF pages 57–65
  • Normality constructed with Excel and EDPD framework: Manuscript Section 3.6.4 and Figure 3-1, PDF pages 66–68
  • Human ICC and comparison across variants: Manuscript Sections 4.3–4.5, PDF pages 76–99; Tables 4-1 to 4-8
  • Human role-play design and deviation analysis: Manuscript Sections 5.2–5.3, PDF pages 102–116; Tables 5-1 to 5-15
  • LLM role-play results and omissions: Manuscript Sections 5.4–5.5, PDF pages 116–120; Tables labeled 5-16 to 5-18
  • Model size and identity errors: Manuscript Sections 3.5.2 and 5.4.2, PDF pages 53–55 and 117; official ChatGLM3 repository, Meta Llama 3.1 model card, and DeepSeek V3/R1 repositories checked 15 July 2026
  • Conclusions and limitations declared by authors: Manuscript Sections 6.1–6.2 and Chapter 7, PDF pages 121–126
  • Absence of later version and artifacts: arXiv API plus exact-title, arXiv-ID, GitHub and web searches checked 15 July 2026
  • Complete visual inspection: All 142 pages of arXiv:2505.14845v1 rendered and visually inspected on 15 July 2026