Large Language Models Demonstrate Distinct Personality Profiles

Evaluation and psychometric validity2025PubMed CentralApproved editorial review

Authors: Thomas F Heston, Justin Gillette

Keywords: AI ethics, AI in mental health, AI psychometrics, Artificial intelligence in medicine, Generative AI, Personality assessment

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

2
Authors
20
Findings
79
Limitations
17
Evidence

Editorial summary

English

The study compares how four conversational products available in April 2024 respond to two human personality questionnaires: ChatGPT-3.5, Gemini Advanced, Claude 3 Opus, and Grok in Regular mode. Each product completes the 32 Open Extended Jungian Type Scales (OEJTS) pairs and the 50 IPIP Big-Five items fifteen times in fresh chats. The authors remove the midpoint, force four response options, and map 1, 2, 3, and 4 onto the original 1, 2, 4, and 5 scoring scale. Prompts are not neutral: they tell the model to “play the role of yourself,” be “100% honest,” and return a mandatory number. Sixty administrations per instrument are analyzed with MANOVA, ANOVA, and frequencies; no clinical conversations or patients are evaluated.

Large product differences appear under this protocol. For OEJTS, the MANOVA reports Wilks' lambda 0.130, F(12, 140.52)=13.63, p<0.001. Claude returns INTJ in 15/15 administrations; ChatGPT's modal type is ENTJ in 7/15, Gemini's INFJ in 9/15, and Grok's INFJ in 8/15. For Big Five, the MANOVA reports lambda 0.115, F(15, 143.95)=11.44, p<0.001, and partial eta squared 0.514. Gemini scores 68.7±12.5 on Agreeableness and 82.6±9.0 on Conscientiousness, below the other three; Claude reaches 97.1±1.0 on Conscientiousness; Grok has the highest Openness, 95.0±1.4, but also the greatest dispersion in Emotional Stability, 81.1±20.4. The defensible result is that scored outputs from these four configurations differ and that some recur with substantial regularity.

Fifteen sessions do not make fifteen people or independent samples from a population of models. They are stochastic replicates of the same product sharing weights, alignment, and interface. The power calculation and inferential tests treat each administration as an independent observation, inflating the effective sample size for model-level inference. Exact snapshot identifiers, temperature, top-p, seed, provider system prompts, account configuration, and execution days are not recorded. Methods announce one nine-variable MANOVA, 95% confidence intervals, and Bonferroni comparisons; Results present two separate MANOVAs, show no intervals, omit the post-hoc table, and do not report F, degrees of freedom, p, and effect size for each ANOVA. Levene's test is also significant for all five Big Five traits, with no robust alternative or complete assumption assessment.

Construct validity is the main limitation. OEJTS and Big Five were developed for human self-report, and the paper acknowledges that they have not been validated in artificial systems. It tests no factor structure, measurement invariance, internal reliability, convergence between instruments, or correspondence with external behavior. Removing neutrality to “increase discrimination” and then concluding that models are non-neutral is circular; no neutral score is defined or tested. Variation across chats does not rule out memorization or prior exposure to questionnaire content, while the instruction to answer “as yourself” induces the very self-referential performance being interpreted. “Personality” should therefore be restricted to a prompt- and interface-conditioned response profile.

The integrity audit finds two clear numerical contradictions. Table 1 says Gemini selected Sensing in 60% of fifteen trials, nine trials, while Table 2 says it produced INFJ, which requires Intuition, in nine of those same fifteen; both cannot be true. The abstract says Claude had the highest Emotional Stability, but Table 3 gives Gemini 94.1 and Claude 93.2. The abstract also uses the 0.115 lambda from the Big Five analysis to summarize typological and dimensional traits together, whereas OEJTS has lambda 0.130. A paragraph on interdisciplinary oversight is duplicated verbatim in the discussion. The linked Zenodo record exists and lists XLSX and SAV files, but the service returned 503 during the audit and prevented inspection; the record lists no separate response transcripts despite the paper promising “raw outputs.” With no clinical task, sustained interaction, therapeutic outcome, or safety test, recommendations about model selection, patient matching, and formal clinical personality evaluation are governance proposals rather than empirical conclusions of this experiment.

Español

El estudio compara la forma en que cuatro productos conversacionales disponibles en abril de 2024 responden a dos cuestionarios humanos de personalidad: ChatGPT-3.5, Gemini Advanced, Claude 3 Opus y Grok en modo Regular. Cada producto completa quince veces las 32 parejas del Open Extended Jungian Type Scales (OEJTS) y los 50 ítems IPIP Big-Five, siempre en chats nuevos. Los autores eliminan el punto medio, fuerzan cuatro opciones y convierten 1, 2, 3 y 4 en 1, 2, 4 y 5 para la puntuación original. Los prompts no son neutrales: piden al modelo «interpretar el papel de ti mismo», ser «100% honesto» y elegir obligatoriamente un número. Se analizan 60 administraciones por instrumento con MANOVA, ANOVA y frecuencias; no se evalúan conversaciones clínicas ni pacientes.

Bajo este protocolo aparecen diferencias grandes entre productos. En OEJTS, el MANOVA informa lambda de Wilks 0,130, F(12, 140,52)=13,63, p<0,001. Claude produce INTJ en 15/15 administraciones; el tipo modal de ChatGPT es ENTJ en 7/15, el de Gemini INFJ en 9/15 y el de Grok INFJ en 8/15. En Big Five, el MANOVA informa lambda 0,115, F(15, 143,95)=11,44, p<0,001 y eta cuadrado parcial 0,514. Gemini obtiene amabilidad 68,7±12,5 y responsabilidad 82,6±9,0, por debajo de los otros tres; Claude alcanza responsabilidad 97,1±1,0; Grok obtiene la mayor apertura, 95,0±1,4, pero también la mayor dispersión en estabilidad emocional, 81,1±20,4. El resultado defendible es que las salidas puntuadas de estas cuatro configuraciones difieren y algunas se repiten con bastante regularidad.

Eso no convierte las quince sesiones en quince personas ni en muestras independientes de una población de modelos. Son réplicas estocásticas del mismo producto, con pesos, alineamiento e interfaz compartidos. El cálculo de potencia y las pruebas inferenciales tratan cada administración como observación independiente, por lo que magnifican el tamaño muestral útil para inferir diferencias entre modelos. Tampoco se registran identificadores exactos de snapshot, temperatura, top-p, semilla, system prompts del proveedor, configuración de cuenta o días de ejecución. La sección de métodos anuncia un MANOVA de nueve variables, descriptivos con intervalos del 95% y comparaciones Bonferroni; los resultados presentan dos MANOVA separados, no muestran los intervalos, omiten la tabla post hoc y no proporcionan F, grados de libertad, p y efecto para cada ANOVA. Además, Levene resulta significativo en los cinco rasgos Big Five y no se ofrece una alternativa robusta ni una evaluación completa de supuestos.

La validez de constructo es el límite principal. OEJTS y Big Five fueron desarrollados para autoinforme humano y el artículo reconoce que no están validados en sistemas artificiales. No se prueban estructura factorial, invariancia, fiabilidad interna, convergencia entre instrumentos o correspondencia con conducta externa. Eliminar la neutralidad para «aumentar discriminación» y después concluir que los modelos no son neutrales introduce circularidad; tampoco se define una puntuación neutral ni se contrasta contra ella. La variación entre chats no descarta memorización o conocimiento de cuestionarios, y la instrucción de responder «como tú mismo» induce precisamente la actuación autorreferencial que se interpreta. Por ello, «personalidad» debe limitarse a un perfil de respuesta condicionado por este prompt y estas interfaces.

La auditoría de integridad encuentra dos contradicciones numéricas claras. Table 1 dice que Gemini eligió Sensing en 60% de 15 ensayos, es decir, nueve; Table 2 dice que produjo INFJ, que exige Intuition, en nueve de esos mismos quince: ambas afirmaciones no pueden ser simultáneamente ciertas. El abstract afirma que Claude obtuvo la mayor estabilidad emocional, pero Table 3 da 94,1 a Gemini y 93,2 a Claude. El abstract usa además la lambda 0,115 del análisis Big Five para resumir conjuntamente rasgos tipológicos y dimensionales, mientras OEJTS tiene lambda 0,130. El párrafo sobre supervisión interdisciplinar aparece duplicado literalmente en la discusión. El registro Zenodo enlazado existe y lista archivos XLSX y SAV, pero durante la auditoría el servicio devolvió 503 y no permitió inspeccionarlos; el registro no lista por separado transcripciones textuales pese a que el artículo promete «raw outputs». En ausencia de tareas clínicas, interacción prolongada, resultados terapéuticos o pruebas de seguridad, las recomendaciones sobre selección de modelos, emparejamiento con pacientes y evaluación clínica formal son propuestas de gobernanza, no conclusiones empíricas del experimento.

Research question

Do ChatGPT-3.5, Gemini Advanced, Claude 3 Opus, and Grok-Regular produce different and repeatable response profiles when completing OEJTS and IPIP Big Five in new chats, and can those profiles inform pre-use clinical evaluation?

Method

In April 2024, two questionnaires divided into blocks were administered fifteen times per product: 32 OEJTS pairs and 50 IPIP Big-Five items. Each administration began in a new chat and used identical prompts that instructed responding as "oneself" on a forced four-point scale; the midpoint was removed and categories were mapped to 1, 2, 4, and 5. The authors calculated scores for four OEJTS dimensions and five Big Five dimensions, modal typologies, descriptives, two reported MANOVAs, and univariate ANOVAs. The audit read and rendered the 15 pages, checked the three tables, all prompts, and 40 references, cross-checked PubMed/PMC and the official JATS, and reviewed the Zenodo dataset record, whose download returned 503.

Sample: The statistical sample consists of 15 repeated administrations per each of four products, 60 per instrument. The observed unit is a stochastic execution in a new session, not a person or an independently trained model. The four products are fixed groups chosen by accessibility and willingness to answer; ChatGPT-4 was excluded for refusing. No report is given on how many accounts, days, regions, snapshots, or sampling configurations participated.

Findings

  • Each product completed fifteen administrations of OEJTS and fifteen of Big Five.
  • The OEJTS MANOVA reports Wilks lambda 0.130, F(12,140,52)=13.63, and p<0.001.
  • Claude 3 Opus was classified INTJ in all fifteen OEJTS administrations.
  • The modal OEJTS type for ChatGPT-3.5 was ENTJ in 7/15 administrations.
  • The reported OEJTS type for Gemini Advanced was INFJ in 9/15 administrations.
  • The modal OEJTS type for Grok-Regular was INFJ in 8/15 administrations.
  • The Big Five MANOVA reports Wilks lambda 0.115, F(15,143,95)=11.44, p<0.001, and partial eta squared 0.514.
  • Gemini had the lowest Big Five extraversion, 42.5±12.0.
  • Gemini had the lowest agreeableness, 68.7±12.5, and conscientiousness, 82.6±9.0.
  • Claude had the highest conscientiousness, 97.1±1.0.
  • Gemini had the highest mean emotional stability, 94.1±3.6, slightly above Claude, 93.2±6.1.
  • Grok had the highest openness or intellect, 95.0±1.4.
  • Grok showed the greatest dispersion in emotional stability, 81.1±20.4.
  • The univariate ANOVAs are described as significant for all five traits with eta squared between 0.193 and 0.738.
  • Levene was significant for all five Big Five traits, indicating different variances across products.
  • The OEJTS and Big Five profiles do not provide a concordant unitary classification.
  • The protocol produces differentiable response profiles among these four configurations.
  • The study does not test any clinical task, therapeutic outcome, or safety harm.
  • Table 1 and Table 2 contain an arithmetic contradiction regarding Gemini Sensing/Intuition.
  • The abstract contradicts Table 3 by attributing the highest emotional stability to Claude.

Limitations

  • Only four conversational products were evaluated.
  • The products were chosen by accessibility and willingness to answer, not by representative sampling.
  • ChatGPT-4 was excluded because it refused to answer, introducing selection by compliance.
  • The exclusion removes precisely a behavior relevant to avoiding anthropomorphization.
  • No exact model or snapshot identifiers are provided.
  • Gemini Advanced and Grok-Regular are product labels, not reproducible versions of the underlying model.
  • Exact days and times of execution within April 2024 are not declared.
  • No region, account, plan, or interface configuration is reported.
  • No system prompts or internal provider policies are documented.
  • No temperature, top-p, seed, or other sampling parameters are documented.
  • It is unknown whether the platforms changed snapshots during the 15 repetitions.
  • Identical user prompts do not imply identical complete conditions across providers.
  • The fifteen sessions are replicates of the same product, not independently trained models.
  • Treating sessions as independent samples produces pseudoreplication for inferences across models.
  • The power calculation uses 60 administrations as if they were independent units.
  • It is not justified that n=15 per product represents temporal or user variability.
  • The study is not replicated with different accounts, dates, regions, or operators.
  • The prompt explicitly instructs interpreting the role of oneself.
  • The prompt instructs being 100% honest and not hallucinating.
  • The conditions are not "unprompted" as the abstract claims.
  • The neutral midpoint is removed from both instruments.
  • The removal of neutrality forces a direction in each response.
  • Concluding absence of neutrality after preventing neutral responses introduces circularity.
  • A neutral profile is not operationally defined.
  • Each model is not tested against a prespecified neutral score.
  • The mapping of 1,2,3,4 to 1,2,4,5 alters intervals and properties of the original instrument.
  • The reference on midpoint use comes from human surveys and does not validate the modification in LLMs.
  • OEJTS was designed for human self-report and is not validated in artificial systems.
  • IPIP Big Five was designed for humans and is not validated in artificial systems.
  • Factorial structure in model responses is not evaluated.
  • Invariance across products is not evaluated.
  • No alpha, omega, or other internal consistency per product is reported.
  • Test-retest reliability with a controlled time interval is not studied.
  • Convergence between OEJTS and Big Five with predefined hypotheses is not evaluated.
  • No validation against external behavior or spontaneous dialogue is performed.
  • There is no blind human evaluation of style or behavior.
  • Answering questionnaires may reflect instruction compliance and learned stereotypes.
  • Sensitivity to prompt paraphrasing is not tested.
  • Sensitivity to item or option order is not tested.
  • Contamination or memorization of the questionnaires is not tested.
  • Variation across administrations does not rule out memorization, contrary to the discussion argument.
  • The methods section announces a MANOVA with nine dependent variables.
  • The results present separately a four-variable OEJTS MANOVA and a five-variable Big Five MANOVA.
  • The combined nine-variable MANOVA described in methods is not reported.
  • The abstract uses lambda 0.115 to jointly summarize typology and dimensions, although OEJTS reports 0.130.
  • Box's M or other covariance homogeneity tests are not reported.
  • No multivariate normality evaluation is reported.
  • The variables are discrete, bounded, and in some cases nearly invariant.
  • Levene is significant for all five Big Five traits.
  • No robust alternative is applied in the presence of heteroscedasticity.
  • The 95% confidence intervals promised in methods are not shown.
  • The promised Bonferroni post hoc comparison table is not published.
  • No specific contrasts explaining each difference across products are indicated.
  • No F, degrees of freedom, p, and effect are reported for each univariate ANOVA.
  • The eta squared range is given without a complete table per trait.
  • No SPSS code, syntax, or results file is provided.
  • The scoring and classification algorithm is not fully explained in reproducible code.
  • Table 1 attributes Sensing to Gemini in 60% of 15 administrations.
  • Table 2 attributes INFJ, which requires Intuition, to Gemini in 9/15 administrations.
  • The two figures for Gemini cannot coexist in a sample of fifteen.
  • The abstract says Claude has higher emotional stability.
  • Table 3 shows higher emotional stability in Gemini, 94.1 versus 93.2.
  • The discussion compares effect sizes with human subgroups without a direct statistical comparison.
  • A paragraph on interdisciplinary supervision is duplicated verbatim on page 7.
  • The Zenodo record only lists one XLSX and one SAV and not a separate transcription of textual responses.
  • Zenodo returned 503 during the audit and did not allow verification of the linked files.
  • Medical or mental health prompts are not evaluated.
  • Multi-turn conversations are not evaluated.
  • No patients, clinicians, or mental health evaluators participate.
  • Therapeutic alliance, trust, satisfaction, or clinical outcome are not measured.
  • Diagnostic errors, triage, crisis, or harmful advice are not evaluated.
  • It is not tested that scores predict clinical safety.
  • Matching between model personality and patient preferences is not tested.
  • The analogy with adapting pharmacotherapy is not supported by the design.
  • Regulatory recommendations exceed an isolated questionnaire experiment.
  • The study is a snapshot from April 2024 and several products no longer represent current systems.
  • Other languages, cultures, or user populations are not studied.
  • Effects of training, alignment, interface, and randomness are not separated.
  • Stability after provider updates is not quantified.

What the study does not establish

  • It does not demonstrate that LLMs have personality, emotions, or human self-concept.
  • It does not identify internal traits independent of prompt and interface.
  • It does not demonstrate neutrality or absence of neutrality by means of a valid test.
  • It does not validate OEJTS or Big Five for artificial systems.
  • It does not demonstrate psychometric invariance across the four products.
  • It does not demonstrate that fifteen chats are independent samples from a population of models.
  • It does not allow attributing differences to architecture rather than alignment, product, or sampling.
  • It does not establish current profiles for current versions of ChatGPT, Gemini, Claude, or Grok.
  • It does not demonstrate that OEJTS typologies are stable outside this prompt.
  • It does not demonstrate that Big Five profiles predict conversational behavior.
  • It does not rule out memorization or familiarity with the items.
  • It does not demonstrate safety, efficacy, or therapeutic adequacy.
  • It does not demonstrate that matching model and patient personality improves outcomes.
  • It does not by itself justify clinical personality evaluations as a regulatory requirement.
  • It does not allow exactly reproducing the experiment without snapshots and generation parameters.

Traceability

Scope: Full text

Version: Cureus 17(5):e84706 (23 May 2025); PMID 40551914; PMCID PMC12183331; DOI 10.7759/cureus.84706; CC BY 4.0

Consulted source: https://pmc.ncbi.nlm.nih.gov/articles/PMC12183331/

Review: Codex full-text, visual, psychometric, statistical-assumption, arithmetic-integrity, clinical-claim and source-transparency audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • ChatGPT-3.5 through the public ChatGPT product in April 2024; exact snapshot not reported
  • Gemini Advanced through the public product in April 2024; exact model identifier not reported
  • Claude 3 Opus through the public product in April 2024; exact snapshot not reported
  • Grok Regular Mode through the public product in April 2024; exact model identifier not reported
  • ChatGPT-4 attempted and excluded after refusing anthropomorphic questionnaire items

Instruments and metrics

  • Open Extended Jungian Type Scales v1.2: 32 paired items and four IE, SN, TF and JP dimensions
  • IPIP Big-Five Factor Markers: 50 items scored as Extraversion, Emotional Stability, Agreeableness, Conscientiousness and Intellect/Openness
  • Modified four-point forced-choice scale with original midpoint removed
  • Mapping from response categories 1, 2, 3, 4 to scores 1, 2, 4, 5
  • Modal OEJTS four-letter classification
  • One-way multivariate analysis of variance
  • Follow-up univariate ANOVA and declared Bonferroni comparisons
  • Wilks' lambda, Pillai's trace, Hotelling's trace and Roy's largest root
  • Partial eta squared and eta squared
  • Levene tests for equality of variance

Data used

  • 60 OEJTS administrations: 15 per conversational product
  • 60 IPIP Big-Five administrations: 15 per conversational product
  • Published aggregate Tables 1–3
  • Zenodo dataset 10.5281/zenodo.11087767 listing one XLSX and one SPSS SAV file
  • Official Europe PMC JATS XML for PMCID PMC12183331

Evidence and location

  • Design, products, and declared clinical objectives: Cureus 17(5):e84706, abstract and introduction, pp. 1–2
  • Product selection and exclusion of ChatGPT-4: Paper, Study population, pp. 2–3
  • Instruments, midpoint removal, and scale mapping: Paper, Intervention, p. 3
  • Fifteen administrations per product and instrument: Paper, Intervention and Sample size calculation, p. 3
  • Statistical plan of nine variables, intervals, and Bonferroni: Paper, Statistical analysis, p. 3
  • MANOVA and OEJTS profiles: Paper, Results and Tables 1–2, pp. 4–5
  • Big Five means and MANOVA: Paper, Table 3 and Results, p. 5
  • Heterogeneity of variances: Paper, final Results paragraph, p. 5
  • Gemini Sensing contradiction versus INFJ: Paper, Tables 1–2, pp. 4–5; arithmetic integrity audit 15 Jul 2026
  • Higher emotional stability contradiction: Paper, abstract p. 1 versus Table 3 p. 5; integrity audit 15 Jul 2026
  • Clinical interpretation and memorization argument: Paper, Discussion, pp. 5–7
  • Limitations acknowledged by the authors: Paper, Limitations, p. 7
  • Complete OEJTS and IPIP prompts: Paper, Appendices, pp. 8–13
  • Publication, DOI, license, and non-retracted status: PubMed PMID 40551914; PMCID PMC12183331; DOI 10.7759/cureus.84706; NCBI OA record; checked 15 Jul 2026
  • Official structured text: .cache/editorial-sources/article-077/supplements/audit/PMC12183331-fulltext.xml; sha256 4c0efb186c61c8e29f426a2ba6c1c8f8eda4fc306b435ca682f299d216ec2bf5
  • Registered dataset and temporary delivery failure: Zenodo 10.5281/zenodo.11087767 lists XLSX md5 11f3266b41a8c01d6e26d19fc2128a47 and SAV md5 d87f2844823a149957d177e879844b40; HTTP 503 during audit 15 Jul 2026
  • Integral reading and visual inspection: All 15 PDF pages rendered and inspected, including Tables 1–3, appendices and 40 references; checked 15 Jul 2026