Investigating Large Language Models in Inferring Personality Traits from User Conversations

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Jianfeng Zhu, Ruoming Jin, Karin G. Coifman

Keywords: Large Language Models, Personality, Big Five, Persona, Personality Control

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

3
Authors
14
Findings
38
Limitations
10
Evidence

Editorial summary

English

This preprint evaluates whether GPT-4o and GPT-4o mini can infer Big Five traits from semi-structured interview transcripts for 102 participants. Each person answered five prompts about the previous day, a challenging experience, coping, an unpleasant event, and a positive event; responses were paired with BFI-10 self-reports and depressive-symptom information. The study compares two zero-shot routes: directly request five trait scores from 1 to 5, or first request all ten BFI-10 item scores and then calculate traits. The sample is highly imbalanced: 96 men and 6 women, ages 24–56, mean 36.63. Text is lowercased and stripped of stopwords and punctuation before model input. Direct results are weak. Pearson correlations with human BFI-10 scores range from −0.090 to 0.185; GPT-4o obtains 0.098 for Extraversion, 0.184 for Agreeableness, −0.058 for Conscientiousness, 0.025 for Neuroticism, and 0.142 for Openness. GPT-4o mini obtains 0.151, −0.063, −0.019, −0.090, and 0.185. Across the ten BFI items no correlation reaches 0.3; the joint range is −0.216 to 0.259. GPT-4o has the smaller absolute mean bias on four items and mini on six, but this is not individual error. The paper calls the signed mean difference between reference and prediction “accuracy.” Positive and negative errors can cancel, so a near-zero mean does not imply participant-level predictions are close. The paper reports no MAE, RMSE, concordance correlation, calibration, intervals, tests, or error dispersion. Reconstructing Table 3, the BFI-10 route reduces absolute mean bias in 7 of 10 model-trait combinations and worsens it in 3: GPT-4o/Extraversion moves from |−0.514| to |0.598|; mini/Conscientiousness from |0.157| to |0.475|; and mini/Openness from |0.162| to |−0.294|. The unweighted mean falls from 0.446 to 0.371, but it remains aggregate bias rather than MAE. Claims of a significant improvement and greater accuracy are therefore not established. The symptom analysis splits participants between none and at least one symptom. The method first attributes labels to SCID/SCID-5 interviews and later says that “a survey like PHQ-9” was used, without identifying the exact instrument or rule. Group sizes are not stated; table increments imply 79 without symptoms and 23 with symptoms, an arithmetic inference rather than a reported fact. Comparing mean bias within each group does not demonstrate sensitivity to a depression-related shift: reference distributions may differ, predicted and observed group deltas are not directly compared, and no uncertainty is reported. Mini’s “perfect alignment” on BFI-4 means only zero mean bias, not participant-level equality. Moreover, for BFI-9 in the symptom-present group Table 5 favors GPT-4o, 1.000 versus 1.087 absolute bias, while the prose says mini is better. Exact model snapshots, execution date, API parameters, seeds, repeat count, parser, exact BFI-10 scoring formula, code, and data are not reported. The faithful conclusion is that both models have very low correlations with the reference and systematic bias in this small, skewed sample. Imposing BFI-10 structure changes and sometimes reduces group-level bias, but it does not establish individual accuracy, reliability, clinical sensitivity, or an ability to assess personality or depression.

Español

Este preprint evalúa si GPT-4o y GPT-4o mini pueden inferir Big Five desde transcripciones de entrevistas semiestructuradas de 102 participantes. Cada persona respondió cinco preguntas sobre el día anterior, una experiencia difícil, afrontamiento, un evento desagradable y uno positivo; las respuestas se unieron a autoinformes BFI-10 y a información sobre síntomas depresivos. El estudio compara dos rutas zero-shot: pedir directamente cinco scores de rasgo entre 1 y 5, o pedir primero los diez ítems BFI-10 y calcular después los rasgos. La muestra está muy desequilibrada: 96 hombres y 6 mujeres, edades 24–56, media 36,63. Antes de enviarlo al modelo, el texto se pasa a minúsculas y se eliminan stopwords y puntuación. Los resultados directos son débiles. Las correlaciones de Pearson con el BFI-10 humano van de −0,090 a 0,185; GPT-4o obtiene 0,098 en extraversión, 0,184 en amabilidad, −0,058 en responsabilidad, 0,025 en neuroticismo y 0,142 en apertura. GPT-4o mini obtiene 0,151, −0,063, −0,019, −0,090 y 0,185. En los diez ítems BFI, ninguna correlación llega a 0,3: el rango conjunto es −0,216 a 0,259. GPT-4o presenta menor sesgo medio absoluto en cuatro ítems y mini en seis, pero esto no es error individual. El paper llama “accuracy” a la diferencia media firmada entre referencia y predicción. Como errores positivos y negativos pueden cancelarse, un promedio cercano a cero no implica predicciones cercanas persona a persona. No se reportan MAE, RMSE, correlación de concordancia, calibración, intervalos, tests ni dispersión de errores. Al reconstruir la Tabla 3, la ruta BFI-10 reduce el valor absoluto de ese sesgo en 7 de 10 combinaciones modelo-rasgo y lo empeora en 3: GPT-4o/extraversión pasa de |−0,514| a |0,598|; mini/responsabilidad de |0,157| a |0,475|; y mini/apertura de |0,162| a |−0,294|. La media no ponderada baja de 0,446 a 0,371, pero sigue siendo sesgo agregado, no MAE. Por ello, las afirmaciones de mejora “significativa” y mayor exactitud no están demostradas. La sección de síntomas divide entre ninguno y al menos uno. El método primero atribuye etiquetas a entrevistas SCID/SCID-5 y después dice que se empleó “una encuesta como PHQ-9”, sin identificar el instrumento exacto ni el criterio. No publica tamaños de grupo; los incrementos de las tablas implican 79 sin síntomas y 23 con síntomas, inferencia aritmética no declarada. Comparar el sesgo medio por grupo no prueba sensibilidad a un cambio depresivo: las distribuciones de referencia pueden diferir y no se contrastan los deltas predichos con los deltas observados, ni hay incertidumbre. “Alineación perfecta” de mini en BFI-4 significa únicamente sesgo medio 0, no igualdad individual. Además, para BFI-9 con síntomas la Tabla 5 muestra menor sesgo en GPT-4o, 1,000 frente a 1,087, pero el texto afirma que mini fue mejor. No se identifican snapshots de los modelos, fecha de ejecución, parámetros de API, seeds, número de repeticiones, parser, fórmula exacta de scoring BFI-10, código ni datos. La conclusión fiel es que, en esta muestra pequeña y sesgada, ambos modelos muestran correlaciones muy bajas con la referencia y sesgos sistemáticos. Imponer el formato BFI-10 modifica y a veces reduce el sesgo de grupo, pero no demuestra exactitud individual, fiabilidad, sensibilidad clínica ni capacidad para evaluar personalidad o depresión.

Research question

With what fidelity do GPT-4o and GPT-4o mini assign Big Five traits and BFI-10 items from five interview responses, whether calculating traits from predicted items improves the comparison with self-report, and whether the bias of the predictions differs between groups with and without depressive symptoms?

Method

Zero-shot evaluation of two models on 102 transcripts. One prompt directly produces five Big Five scores; another produces ten BFI-10 scores that are then converted to traits using an unpublished formula. Predictions are compared with human BFI-10 using Pearson and signed mean difference. It is repeated descriptively by groups of zero versus at least one depressive symptom, without tests, intervals, or individual error metrics.

Sample: 102 participants: 96 men and 6 women; 24–56 years, mean 36.63. Each one answers the same five questions and provides BFI-10. The paper does not report sizes of the symptom groups; the fractions of Tables 4–5 are compatible with 79 without symptoms and 23 with at least one. Recruitment, clinical composition, race, education, language, consent, or ethical protocol are not described.

Findings

  • The current source is arXiv:2501.07532v1, submitted on 13 January 2025, with 13 pages.
  • The 13 pages were rendered and visually inspected.
  • The current PDF matches byte for byte with the cache and has SHA-256 897fa2ad6837ececb2191d4bc5beda628e9a6748fcc2096a5f2ad4be88a12fd8.
  • The direct model-BFI correlations per trait range from −0.090 to 0.185; none reaches 0.2.
  • The per-item BFI correlations range from −0.216 to 0.259; none reaches 0.3.
  • GPT-4o has a lower absolute mean difference value in 4 of 10 items and GPT-4o mini in 6 of 10.
  • The BFI-10 route reduces the absolute value of the mean bias in 7 of 10 model-trait combinations and increases it in 3.
  • The unweighted average of the absolute mean bias goes from 0.4464 in direct prediction to 0.3711 in the BFI-10 route.
  • GPT-4o/extraversion worsens from 0.514 to 0.598 absolute bias with the BFI-10 route.
  • GPT-4o mini/conscientiousness worsens from 0.157 to 0.475 and mini/openness from 0.162 to 0.294.
  • The mean difference of mini on BFI-4 for the group with symptoms is 0, but this only indicates net cancellation or absence of mean bias.
  • For BFI-9 with symptoms, Table 5 favors GPT-4o with 1.000 versus 1.087 of mini, contrary to the text.
  • The denominators of the tables imply most likely 79 participants without symptoms and 23 with symptoms, although the paper does not declare them.
  • No code, data, predictions, raw responses, reproducible configuration, or peer-reviewed publication were located.

Limitations

  • The mean difference is signed and allows cancellation of individual errors.
  • The paper describes it incorrectly as a measure of absolute accuracy.
  • No MAE, median absolute error, RMSE, residual distribution, or percentage within a tolerance is reported.
  • A zero mean bias does not imply perfect person-by-person alignment.
  • The Pearson correlations are very low and are not accompanied by intervals or p-values.
  • Multiplicity is not corrected despite comparing five traits, ten items, two models, two routes, and two groups.
  • “Significantly” is used without hypothesis testing or estimation of uncertainty.
  • The BFI-10 improvement is not uniform: three of ten combinations worsen in absolute bias.
  • The unweighted aggregate comparison treats traits and models as equivalent and does not use per-participant error.
  • The exact formula used to convert the ten predicted items to five traits is not published.
  • The reverse scoring of the five negative items and the final range of the trait are not clarified.
  • The BFI prompt first asks for a detailed explanation and then prohibits any explanation, a contradictory instruction.
  • No parser, treatment of invalid outputs, decimals, or out-of-range values is described.
  • The exact snapshot of GPT-4o or GPT-4o mini is not identified.
  • No execution date, temperature, top-p, seed, token limit, system prompt, or API version is reported.
  • The queries are not repeated, so stochastic variability is unknown.
  • The sample of 102 is small for group differences and five trait correlations.
  • The sample has 94% men and only six women, which limits generalization.
  • No race, ethnicity, education, language, country, recruitment, or full clinical context is reported.
  • Removing stopwords can delete negations and function words relevant to personality.
  • Removing punctuation deletes pauses, emphasis, and discursive signals from the interviews.
  • Converting to lowercase and filtering text reduces the intended validity of real conversation.
  • There is no comparison with simple baselines such as training mean, lexical regression, or embeddings.
  • There is no development split because the study is zero-shot, but neither is there external validation in another sample.
  • The BFI-10 has only two items per trait and is a brief reference with measurement error not quantified here.
  • Inference from interview and BFI-10 self-report are not the same mode of measurement.
  • The direct prompt delivers construct descriptors and can induce lexical coincidence.
  • SCID/SCID-5 and a survey such as PHQ-9 are mentioned as the source of symptoms without reconciliation.
  • The threshold of at least one symptom does not equate to a diagnosis of depression, a fact partially recognized by the paper.
  • Sizes and severity distribution of the groups are not explicitly reported.
  • The 79/23 inference comes from the decimal denominators of the tables and does not replace the methodological report.
  • Comparing bias by group does not prove that the model captures the difference between groups.
  • Predicted deltas are not directly compared against observed BFI deltas.
  • Age, gender, or other confounders between groups are not controlled.
  • The BFI-9 contradiction shows that at least one narrative interpretation does not follow the table.
  • No consent, IRB, privacy management, or conditions for reuse of sensitive interviews are reported.
  • There is no code, anonymized data, input hashes, or per-participant results to reproduce the analysis.
  • The work is a preprint v1 without located evidence of peer review.

What the study does not establish

  • It does not establish exact prediction of Big Five at the individual level.
  • It does not demonstrate a statistically significant improvement with the BFI-10 step.
  • It does not demonstrate that low mean bias implies low individual error.
  • It does not demonstrate reliability across runs or model versions.
  • It does not demonstrate sensitivity to depression or to clinical changes.
  • It does not demonstrate diagnosis of depression or of any disorder.
  • It does not demonstrate generalization to women, other ages, languages, or populations.
  • It does not demonstrate that GPT-4o mini is globally superior to GPT-4o.
  • It does not justify automated psychological assessment, persuasive personalization, or clinical intervention.
  • It does not allow reproducing results without configuration, code, and data.

Traceability

Scope: Full text

Version: arXiv:2501.07532v1, submitted 13 January 2025, 13 pages

Consulted source: https://arxiv.org/abs/2501.07532

Review: Codex complete bilingual full-text fidelity pass, current arXiv-version and byte-level PDF check, all-page visual inspection, signed-bias reconstruction, table-versus-prose consistency audit, psychometric scoring audit, depressive-group denominator and validity audit, reproducibility and artifact-availability assessment; summaries written from the full paper and tables rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o through an unspecified OpenAI ChatGPT/API snapshot
  • GPT-4o mini through an unspecified OpenAI ChatGPT/API snapshot

Instruments and metrics

  • Big Five Inventory-10 self-report as personality reference
  • Direct five-trait 1–5 prompt with one-decimal scores
  • Ten-item BFI-10 scoring prompt
  • Pearson correlation
  • Signed mean difference mislabeled as absolute prediction accuracy
  • Depressive-symptom grouping attributed inconsistently to SCID/SCID-5 and a survey like PHQ-9

Data used

  • Previously collected recorded semi-structured interviews from 102 participants
  • Five interview responses per participant, approximately 15 minutes of speech in total
  • Participant BFI-10 scores
  • Depressive symptom information from an incompletely specified clinical or survey source
  • No public dataset, code repository, predictions, model responses or analysis scripts located

Evidence and location

  • Version, authors, date, length, and abstract: arXiv:2501.07532v1 metadata and title page checked 15 July 2026
  • Sample, interview, BFI-10, symptoms, and preprocessing: Materials and Methods, pages 2–3
  • Direct and BFI-10 prompts: Methods, pages 3–5
  • Definition of metrics: Performance Evaluation Approach, pages 5–6
  • Direct correlations and biases: Table 1 and Figure 1, page 6
  • Per-item results: Table 2 and Figure 2, page 7
  • Direct comparison versus BFI-10 route: Table 3 and Figure 3, pages 7–8; arithmetic reconstruction in reports/verification/article-183-bias-and-reporting-audit.json
  • Per-group results and BFI-9 contradiction: Tables 4–5 and Figures 4–5, pages 8–10
  • Complete visual inspection: All 13 pages of arXiv:2501.07532v1 rendered and visually inspected on 15 July 2026
  • Absence of artifacts: Paper, exact-title, arXiv-ID, GitHub, Hugging Face and Papers With Code checks on 15 July 2026