Large language models can infer psychological dispositions of social media users

Evaluation and psychometric validity2024PNAS NexusApproved editorial review

Original title: Large Language Models Can Infer Psychological Dispositions of Social Media Users

Authors: Heinrich Peters, Sandra C Matz

Keywords: Large Language Models, Personality inference, Big Five, Social media, Zero-shot learning, GPT-3.5, GPT-4

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

2
Authors
11
Findings
37
Limitations
11
Evidence

Editorial summary

English

The paper evaluates whether two early ChatGPT versions can rank Facebook users by Big Five traits from text that was not labeled for that task. It randomly selects 1,000 adult MyPersonality participants who completed the 100-item IPIP and posted at least 200 status updates; mean age is 24.2 and 63.1% are women. Each person's 200 most recent updates are concatenated into ten chunks of 20. GPT-3.5 Turbo 0301 and GPT-4 0314 receive the prompt “Rate the text on the Big Five personality dimensions,” return five scores from 1 to 5, and repeat every inference three times. The authors average repetitions and chunks, then compare the result with self-reports. GPT-3.5 correlations range from r = 0.223 to 0.298 and GPT-4 correlations from 0.264 to 0.327; their means are 0.271 and 0.307, respectively, and 0.289 across all ten results. This corresponds to roughly 5–11% shared variance per trait: there is a modest ranking signal, not precise recovery of an individual's personality. GPT-4 is numerically higher on all five traits, but none of its differences from GPT-3.5 is statistically significant. The score scales are also poorly calibrated. GPT-3.5 underestimates Conscientiousness by 1.254 points and Agreeableness by 1.032 on a 1–5 scale; GPT-4 underestimates Conscientiousness by 0.690 and overestimates Extraversion by 0.881. Correlation therefore does not make model outputs interchangeable with IPIP scores. Twenty status updates already yield correlations of 0.176–0.257; using 200 raises them to 0.222–0.327, with different gains by trait. The demographic analyses compare group means and absolute residuals. For both models, absolute errors are lower for women on Conscientiousness, Agreeableness, and Neuroticism; the Openness difference appears only for GPT-3.5, while GPT-4 makes larger Extraversion errors for women. Within-gender correlations do not differ significantly. By age, GPT-3.5 has larger errors for older users on Openness and Conscientiousness but smaller errors on Agreeableness; GPT-4 shows no significant absolute-error differences. The headline of greater accuracy for women and younger people therefore applies to selected traits and metrics, not a general pattern. Among 68 people with friend ratings, observer–self-report correlations average 0.304 and observer–LLM correlations 0.269–0.276, but equivalence is never formally tested. Table S6.1 has a verifiable reporting error: its confidence intervals and p-values are identical to the GPT-4 rows in S6.2 and do not correspond to the observer–self-report coefficients. It cannot support a strong claim of human parity or shared cue use. The study shows that these snapshots capture textual regularities associated with Big Five self-reports in a highly selected sample. It does not validate a personality test, clinical inference, or a tool suitable for individual decisions. The clearest implication is a privacy risk: even a modest signal can be exploited at scale without users' knowledge. Current reproducibility is inadequate. The paper points to OSF hq2ra, but that project now returns HTTP 401 without authentication; an archived page shows that it was public and stored 6.4 MB in November 2024, while the arXiv source contains only the manuscript and figures, not data or analytical code.

Español

El artículo evalúa si dos versiones tempranas de ChatGPT pueden ordenar a usuarios de Facebook según sus rasgos Big Five a partir de texto no etiquetado para esa tarea. Selecciona aleatoriamente 1.000 participantes adultos de MyPersonality que completaron los 100 ítems del IPIP y publicaron al menos 200 estados; la muestra tiene una edad media de 24,2 años y un 63,1 % de mujeres. Para cada persona se concatenan los 200 estados más recientes en diez bloques de 20. GPT-3.5 Turbo 0301 y GPT-4 0314 reciben el prompt “Rate the text on the Big Five personality dimensions”, devuelven cinco puntuaciones de 1 a 5 y repiten cada inferencia tres veces. Los autores promedian repeticiones y bloques, y comparan el resultado con el autoinforme. Las correlaciones de GPT-3.5 van de r = 0,223 a 0,298 y las de GPT-4 de 0,264 a 0,327; las medias son 0,271 y 0,307, respectivamente, y 0,289 al combinar los diez resultados. Esto representa aproximadamente un 5–11 % de varianza compartida por rasgo: existe una señal de ordenación modesta, pero no una lectura precisa de la personalidad individual. GPT-4 obtiene valores numéricamente superiores en los cinco rasgos, aunque ninguna diferencia frente a GPT-3.5 es estadísticamente significativa. Además, las escalas están mal calibradas: GPT-3.5 subestima en promedio conciencia por 1,254 puntos y amabilidad por 1,032 en una escala de 1 a 5; GPT-4 subestima conciencia por 0,690 y sobreestima extraversión por 0,881. Por tanto, la correlación no convierte las salidas en puntuaciones psicométricas intercambiables con el IPIP. Con 20 estados ya aparecen correlaciones de 0,176–0,257; usar 200 las eleva a 0,222–0,327, con beneficios distintos por rasgo. Los análisis sociodemográficos comparan medias y residuos absolutos. Para ambos modelos, los errores absolutos son menores en mujeres en conciencia, amabilidad y neuroticismo; en apertura la diferencia solo aparece en GPT-3.5 y, en extraversión, GPT-4 comete más error en mujeres. Las correlaciones dentro de cada género no difieren significativamente. Por edad, GPT-3.5 tiene más error en personas mayores para apertura y conciencia, pero menos para amabilidad; GPT-4 no muestra diferencias significativas de error absoluto. Así, el titular de mayor precisión para mujeres y jóvenes describe algunos rasgos y métricas, no un patrón general. En 68 personas con evaluaciones de amistades, las correlaciones observador–autoinforme promedian 0,304 y las de observador–LLM 0,269–0,276, pero no se contrasta formalmente su equivalencia. La tabla S6.1 contiene un error verificable: sus intervalos y valores p son idénticos a los de GPT-4 en S6.2 y no corresponden a las correlaciones observador–autoinforme. Por ello, no se sostiene una afirmación fuerte de paridad humana ni de uso de los mismos indicios. El estudio demuestra que estos snapshots captan regularidades textuales asociadas con autoinformes Big Five en una muestra muy seleccionada; no valida un test de personalidad, una inferencia clínica ni una herramienta apta para decisiones individuales. El riesgo más sólido es de privacidad: incluso una señal modesta puede explotarse a escala sin conocimiento del usuario. La reproducibilidad actual es insuficiente: el artículo remite a OSF hq2ra, pero el proyecto devuelve HTTP 401 sin autenticación; una copia archivada muestra que era público y ocupaba 6,4 MB en noviembre de 2024, mientras la fuente arXiv solo contiene manuscrito y figuras, no datos ni código analítico.

Research question

Can GPT-3.5 and GPT-4 infer in zero-shot the relative positions of users on the five Big Five traits from their Facebook statuses, how does the association change with text volume, and how do errors vary by gender and age?

Method

Correlational study with 1,000 MyPersonality users. The 200 most recent statuses of each user are divided into ten blocks of 20 and sent, three times per inference, to GPT-3.5 Turbo 0301 and GPT-4 0314 using a zero-shot prompt that requests five scores from 1 to 5. The scores are averaged and correlated with the 100-item IPIP. Cumulative curves from 20 to 200 messages are examined, along with t tests and standardized differences in scores and residuals by gender and by age dichotomized at the median, within-group correlations, and an exploratory analysis with friend assessments for 68 people.

Sample: Random sample of 1,000 adults within the MyPersonality subset that completed the 100 IPIP items and had at least 200 Facebook statuses. Mean age 24.2 years, SD 8.8; 63.1% women. Each user contributes their 200 most recent posts, with an average of 17.10 words per status. Observer analyses are limited to 68 people with friend assessments. The exact threshold of the median age and the sizes of its two groups are not reported.

Findings

  • Correlations with self-report are 0.282, 0.223, 0.291, 0.298, and 0.263 for GPT-3.5, and 0.327, 0.264, 0.324, 0.325, and 0.294 for GPT-4 in openness, conscientiousness, extraversion, agreeableness, and neuroticism.
  • The means of r are 0.271 for GPT-3.5 and 0.307 for GPT-4; by trait, r² ranges approximately between 0.05 and 0.11, so that most of the individual variance remains unexplained.
  • GPT-4 is numerically superior across the five traits, but the article reports that no individual difference between models is statistically significant.
  • The scores are not calibrated like the IPIP: GPT-3.5 underestimates conscientiousness and agreeableness in particular, while GPT-4 underestimates conscientiousness and overestimates extraversion.
  • The correlation across the three scoring rounds ranges from 0.88 to 0.96 for GPT-3.5 and from 0.73 to 0.94 for GPT-4, although the article does not define whether these are mean paired correlations nor does it measure absolute agreement.
  • With 20 messages, correlations are already between 0.176 and 0.257; with 200 they reach 0.222–0.327. The gain is not monotonic at every step nor equal across traits.
  • In both models, absolute residuals are smaller for women in conscientiousness, agreeableness, and neuroticism; openness shows this difference only in GPT-3.5, and extraversion shows greater error for women in GPT-4.
  • Correlations computed within gender or age groups do not differ significantly, and standardizing within group does not significantly change the overall correlations.
  • By age, only GPT-3.5 shows greater absolute error in older individuals for openness and conscientiousness, and lower error for agreeableness; GPT-4 shows no significant differences in absolute error.
  • In the subset of 68, mean correlations with friend assessments are 0.304 for self-report, 0.269 for GPT-3.5, and 0.276 for GPT-4, but it is not tested whether these means are equivalent or different.
  • Source verification confirms the published PDF and supplement and the arXiv source package, but not the data or code: OSF hq2ra was public in a November 2024 capture and now requires authentication.

Limitations

  • Correlations of 0.22–0.33 are modest and explain only approximately 5–11% of the variance per trait; they do not justify precise inference at the individual level.
  • Correlation assesses ordering, not calibration. The large differences in mean and variance relative to the IPIP preclude directly interpreting an LLM score as a psychometric score.
  • The IPIP self-report is treated as the criterion, but it also contains measurement error, response bias, and situational variance; the study does not report the instrument's reliability in this sample.
  • The work is zero-shot with respect to the prompt and does not use the sample's labels to fit the model, but it cannot demonstrate that MyPersonality texts, derived posts, or personality tasks did not appear in the training data.
  • The claim that the models were not explicitly trained for the task is not verifiable in proprietary models with unpublished full corpora and objectives.
  • The sample requires at least 200 statuses and voluntary participation in MyPersonality, thereby selecting highly active users interested in psychological tests.
  • The sample is young, with a mean of 24.2 years, and predominantly female; it does not represent the general population nor necessarily current social media users.
  • The texts come from Facebook between 2007 and 2012; language, platform, audience, and posting norms have changed substantially.
  • Only Facebook posts and one prompt in English are evaluated; there is no evidence for other platforms, languages, cultures, or text genres.
  • Two hundred statuses amount on average to about 3,420 words per person, a considerable textual footprint that should not be confused with inference from a few sentences.
  • Statuses are grouped into blocks of 20 and their scores are averaged with equal weight despite wide variation in length; alternatives for aggregation by tokens or messages are not studied.
  • The volume curve accumulates blocks from the 200 most recent statuses, but does not report randomization of order; volume, period, and additional content may change jointly.
  • Only one prompt, a generic system role, and a 1-to-5 scale are used; robustness to formulation, trait order, anchors, examples, or format is not evaluated.
  • Temperature, top-p, exact date of calls, output limits, API errors, and rules for responses that did not meet the format are not reported.
  • Stability across three rounds is summarized with correlations, which do not detect systematic level differences and do not replace an ICC or a measure of absolute agreement.
  • Arithmetically averaging correlations across traits does not use Fisher's transformation and does not convert five distinct constructs into a single accuracy measure.
  • The test used to compare the dependent correlations of GPT-3.5 and GPT-4 on the same users is not described, nor are the statistics or intervals of those differences reported.
  • The comparison with supervised models uses results from another study, with potentially different protocols, and the cited mean of 0.37 is higher than the LLM mean of 0.29; it is not a direct test of similar performance.
  • The gender and age analyses show 80 p-value entries, 70 distinct comparisons after discounting duplicated self-reports, without an explicit correction for multiplicity; several marginal results would not survive a strict correction.
  • The p columns of Tables S4 and S5 print all values with a negative sign, which is mathematically impossible; the main text appears to use their magnitudes, but the error reduces the reliability of the supplement.
  • The study uses binary gender categories and does not explain how this variable was obtained nor does it represent non-binary or trans identities.
  • Age is split at the median, discarding continuous information; the cutoff point and the N of each group are not reported, and the meaning of “young” and “older” depends on a sample with a low age range.
  • Absolute residuals depend on the decalibration of each model and trait; a difference in residuals does not by itself equate to bias in ordering ability.
  • Inferred mean differences may reflect stereotypes, real self-report differences, expression differences, or decalibration. The design does not identify which of these sources produces them.
  • The article claims that there are no differences in correlations within groups, but does not publish the correlations, intervals, N, or tests needed to audit that claim.
  • The observer benchmark has only 68 people, so estimates per five traits are imprecise and sensitive to a few cases.
  • Friendships use a 10-item IPIP, only two per trait, while the self-report uses 100; their lower reliability and instrument differences make comparison difficult.
  • Observers know the person and have more cues than the statuses analyzed by the LLM; they do not perform the same task with the same information.
  • Table S6.1 assigns to the observer–self-report correlations exactly the intervals and p-values of GPT-4 in S6.2. For example, r = 0.198 appears with CI [0.179, 0.583], an interval that does not even contain the coefficient.
  • No equivalence or difference tests are performed between observers and the LLM; the fact that coefficients overlap in range does not demonstrate human parity.
  • The correlation between LLM judgments and observers does not demonstrate that they use the same linguistic cues; that claim would require cue analyses or interventions.
  • The study does not identify which words, topics, styles, or personal data drive the predictions, so it does not distinguish trait, demographics, topic, and stereotype.
  • The OSF repository cited as the source of data and code is not accessible without authentication in the current verification; the analysis cannot be reproduced from the available public artifacts.
  • The arXiv source contains TeX and figures, but no data, API responses, or analytical scripts. Furthermore, its supplement v2 does not even include the S6 analysis added in the final publication.
  • There is no preregistration or declared separation between confirmatory and exploratory analyses, especially for the numerous sociodemographic comparisons.
  • The March 2023 snapshots are withdrawn and cannot be consulted today via the original API; exact replication depends on preserved outputs that are not available.
  • Although the study obtained IRB approval and users donated data to research, applying the finding to non-consented profiles poses a different risk that the experiment does not technically resolve.

What the study does not establish

  • It does not validate the outputs of GPT-3.5 or GPT-4 as a Big Five test equivalent to the IPIP or as a measure suitable for individual decisions.
  • It does not demonstrate that personality, mental health, risk, job suitability, or response to an intervention can be diagnosed from social media posts.
  • It does not prove human understanding or internal psychological representation; a correlation may arise from lexical, thematic, and demographic regularities.
  • It does not demonstrate that the models never saw MyPersonality or derived materials during training.
  • It does not establish that GPT-4 is better than GPT-3.5, because the numerical differences were not significant.
  • It does not demonstrate performance equal to supervised models; the comparison comes from another study and numerically favors the supervised benchmark.
  • It does not demonstrate parity with human observers nor that humans and the LLM use the same cues.
  • It does not establish a general bias against men or older people: patterns change by trait, model, and metric, and within-group correlations do not differ.
  • It does not identify whether demographic differences stem from model training, stereotypes, the sample, or real differences in expression on Facebook.
  • It does not demonstrate that women or young people reveal their personality more faithfully on the internet intrinsically.
  • It does not generalize to current models, other languages, new platforms, short texts, private conversations, or clinical populations.
  • It does not prove that increasing context improves accuracy indefinitely; the curve flattens and fluctuates by trait.
  • It does not demonstrate that the use of automated psychological profiling is ethical, consented, safe, or legal; the article itself underscores privacy and manipulation risks.

Traceability

Scope: Full text

Version: PNAS Nexus 3(6), pgae231; advance access 13 June 2024; PMC11211928.1; CC BY 4.0

Consulted source: https://pmc-oa-opendata.s3.amazonaws.com/PMC11211928.1/PMC11211928.1.pdf

Review: Codex full-text, visual, supplementary-table, arXiv-source and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5 Turbo (gpt-3.5-turbo-0301)
  • GPT-4 (gpt-4-0314)

Instruments and metrics

  • International Personality Item Pool de 100 ítems
  • Big Five: apertura, conciencia, extraversión, amabilidad y neuroticismo
  • Versión IPIP de 10 ítems para evaluaciones de amistades
  • Correlación de Pearson con intervalos del 95 %
  • Correlación entre rondas de puntuación
  • Pruebas t entre grupos
  • Diferencias estandarizadas d
  • Residuos dirigidos y absolutos

Data used

  • MyPersonality Facebook dataset, actividad entre 2007 y 2012
  • 200 estados de Facebook por cada uno de 1.000 usuarios
  • Autoinformes IPIP Big Five de 100 ítems
  • Puntuaciones generadas por GPT-3.5 Turbo 0301 y GPT-4 0314
  • Evaluaciones de personalidad realizadas por amistades para 68 usuarios
  • Suplemento publicado pgae231_supplementary_data.pdf
  • Fuente arXiv 2309.08631v2
  • OSF hq2ra, declarado en el artículo pero no accesible públicamente en la verificación actual

Evidence and location

  • Question, general result, implications, and license: Article, p. 1, Abstract, Significance Statement, and CC BY 4.0 notice
  • Sample, MyPersonality, and IRB: Article, pp. 2–3, Methods, Data and sampling
  • IPIP, snapshots, prompt, blocks, and three repetitions: Article, p. 2, Measures
  • Distributions and decalibration: Article, pp. 2–3, Figure 1; supplement, p. 2, Table S1
  • Correlations by trait and comparison between models: Article, pp. 3–4, Figure 2; supplement, pp. 2–3, Tables S2–S3
  • Differences and residuals by gender: Article, pp. 3–6, Figures 3–4; supplement, p. 4, Table S4
  • Differences and residuals by age: Article, pp. 4–6, Figures 3–4; supplement, p. 5, Table S5
  • Benchmark of 68 observers and interval error: Article, pp. 4–5, Agreement with third-person observer ratings; supplement, p. 6, Tables S6.1–S6.2
  • External comparison with supervised model, limits, and privacy: Article, pp. 5–7, Discussion, Limitations, Implications and Conclusion
  • Sign errors in p-values and content of the previous version: Published supplement, Tables S4–S5; arXiv source 2309.08631v2, SI.tex
  • Current availability of artifacts: OSF hq2ra returns HTTP 401 on 15 July 2026; Wayback 20241125044504 shows the public project with 6.4 MB; the arXiv package contains TeX and figures