Mind Reading or Misreading? LLMs on the Big Five Personality Test

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Francesco Di Cursi, Chiara Boldrini, Marco Conti, Andrea Passarella

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

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

4
Authors
22
Findings
47
Limitations
11
Evidence

Editorial summary

English

This preprint evaluates binary automatic prediction of the five Big Five traits from text with GPT-4 and four local models, Phi-3, Gemma 7B, Llama 3.1 8B, and Mistral 7B, on Essays, MyPersonality, and Pandora. It runs 135 configurations: 75 with a simple prompt across all five models and 60 with an enriched prompt for the four open models only. Each model classifies each trait separately as 0 or 1 at temperature 0 and a 20-token maximum. The enriched prompt supplies Costa and McCrae definitions, positive and negative IPIP phrases, and long Goldberg adjective lists. The paper reports accuracy, macro-F1, and class-specific precision, recall, and F1, together with invalid outputs. The most useful result is diagnostic rather than strong performance. Only 13 of 135 configurations, 9.6%, reach F1 of at least 0.5 for both classes. The strongest balanced row is Mistral on Agreeableness in MyPersonality with the complex prompt: accuracy 0.608, negative-class F1 0.600, positive-class F1 0.620, and macro-F1 0.610, only 7.2 accuracy points above the majority class. GPT-4 on Agreeableness in Essays reaches accuracy 0.593 and macro-F1 0.585; Mistral on Openness in Essays reaches 0.584 and 0.575. Many rows with superficially acceptable accuracy predict the positive class almost exclusively. Enriched prompts usually remove formatting failures and raise positive recall, but often reduce negative recall; they do not uniformly improve accuracy or balance. Extraversion and Neuroticism are the weakest dimensions under these conditions. Interpretation needs additional caution. Invalid outputs are discarded before metrics are calculated, so performance is conditional on model compliance. Phi-3 loses roughly 80–85% of Essays under the simple prompt; its remaining metrics describe a highly selected subset. The F1 ≥ 0.5 cutoff is repeatedly called a significance threshold even though there is no test, null distribution, interval, or multiplicity correction, and the authors acknowledge that 0.5 is not a chance threshold under imbalance. GPT-4 receives a different prompt, is tested only in the simple condition, and has no identified snapshot; claims that complex-prompt open models match it confound model and prompt. GPT-4 is also told that 1 means high or moderate trait level, while the other models classify trait presence or expression. Big Five dimensions are continuous rather than present-or-absent properties, and author-level labels do not guarantee that one text expresses a trait. Pandora adds an unvalidated transformation: its 1–100 scores are interpolated to 1–5 and thresholded using the highest MyPersonality score still labeled 0; the five thresholds are not released. Its 14,221 texts come from 1,608 users, but metrics treat texts as units without author-clustered uncertainty. Internal inconsistencies also matter. For GPT-4/Openness/Pandora, Table 2 reports 679 invalids, while the Table 3 diagonal and Table 4 supports of 7,532+6,209 imply exactly 480; an invalid union cannot be smaller than 679 invalids for one trait. MyPersonality reports 255 texts, but each trait's classes total 250 without explaining five missing labels. Table 1 labels median and SD while the prose says mean and SD. No code, predictions, processed labels, or logs were located. Ollama tags are mutable and lack digests, dates, quantization, and runtime version; GPT-4 lacks a snapshot and date. The defensible conclusion is therefore that zero-shot classification is usually weak or asymmetric in these prompts and datasets, and prompt enrichment changes compliance and class bias. The paper usefully warns about class-specific metrics and abstentions, but it does not establish reliable prediction, statistical significance, a controlled parity comparison with GPT-4, or validity for profiling people.

Español

Este preprint evalúa predicción automática binaria de los cinco rasgos Big Five desde texto mediante GPT-4 y cuatro modelos locales, Phi-3, Gemma 7B, Llama 3.1 8B y Mistral 7B, sobre Essays, MyPersonality y Pandora. Realiza 135 configuraciones: 75 con prompt simple para los cinco modelos y 60 con prompt enriquecido solo para los cuatro modelos abiertos. Cada modelo clasifica por separado cada rasgo con 0 o 1; temperatura 0 y 20 tokens máximos. El prompt enriquecido incorpora definiciones de Costa y McCrae, frases IPIP positivas y negativas y largas listas de adjetivos de Goldberg. Se reportan accuracy, macro-F1 y precision, recall y F1 por clase, además de salidas inválidas. El hallazgo más útil es diagnóstico, no de rendimiento. Solo 13 de 135 configuraciones, 9,6%, alcanzan F1 de al menos 0,5 en ambas clases. La mejor fila equilibrada es Mistral con Agreeableness en MyPersonality y prompt complejo: accuracy 0,608, F1 negativa 0,600, F1 positiva 0,620 y macro-F1 0,610, apenas 7,2 puntos de accuracy sobre la clase mayoritaria. GPT-4 con Agreeableness en Essays obtiene 0,593 de accuracy y macro-F1 0,585; Mistral con Openness en Essays, 0,584 y 0,575. Muchos resultados con accuracy aparentemente aceptable predicen casi siempre la clase positiva. Los prompts enriquecidos suelen eliminar fallos de formato y aumentar recall positivo, pero a menudo reducen recall negativo; no mejoran de forma uniforme accuracy o equilibrio. Extraversion y Neuroticism son las dimensiones más débiles en estas condiciones. La interpretación requiere cautela adicional. Las salidas inválidas se eliminan antes de calcular métricas, por lo que el rendimiento es condicional a que el modelo obedezca. Phi-3 pierde aproximadamente 80–85% de Essays con el prompt simple; sus métricas restantes describen una submuestra muy seleccionada. El corte F1 ≥ 0,5 se denomina repetidamente umbral de significación aunque no hay test, distribución nula, intervalo o corrección múltiple, y los propios autores admiten que 0,5 no equivale a azar bajo desbalance. GPT-4 usa un prompt distinto, solo se prueba en la condición simple y no se identifica su snapshot; afirmar que modelos abiertos con prompt complejo lo igualan mezcla modelo y prompt. Además, GPT-4 recibe la regla 1 para rasgo alto o moderado, mientras los otros reciben presencia o expresión del rasgo. Los Big Five son continuos, no propiedades presentes o ausentes, y las etiquetas de autor no garantizan que un texto concreto exprese el rasgo. Pandora añade otra transformación no validada: sus scores 1–100 se interpolan a 1–5 y se cortan con el mayor valor de MyPersonality aún etiquetado 0; no se publican los cinco umbrales. Sus 14.221 textos proceden de 1.608 usuarios, pero las métricas tratan textos como unidades sin incertidumbre agrupada por autor. Hay también inconsistencias internas. En GPT-4/Openness/Pandora, Tabla 2 reporta 679 inválidos, mientras la diagonal de Tabla 3 y los soportes 7.532+6.209 de Tabla 4 implican exactamente 480; una unión de inválidos no puede ser menor que los 679 de un rasgo. MyPersonality declara 255 textos, pero las clases de cada rasgo suman 250 sin explicar cinco etiquetas ausentes. Tabla 1 rotula mediana y SD, aunque el texto dice media y SD. No se localizaron código, predicciones, etiquetas procesadas o logs. Las tags Ollama son mutables y carecen de digest, fecha, cuantización y versión; GPT-4 carece de snapshot y fecha. Por tanto, la conclusión defendible es que, en estos prompts y datasets, la clasificación zero-shot suele ser débil o asimétrica y el enriquecimiento cambia cumplimiento y sesgo de clase. El trabajo aporta una buena advertencia sobre métricas por clase y abstenciones, pero no demuestra predicción fiable, significación estadística, paridad controlada con GPT-4 ni validez para perfilar personas.

Research question

With what reliability do five LLMs classify binary presence or absence of Big Five traits from texts in three domains, and how do their compliance, class balance, and performance change when enriching the prompt with psychological descriptors?

Method

Zero-shot evaluation across five traits and three datasets. GPT-4, phi3:latest, gemma:7b-instruct, llama3.1:8b, and mistral:instruct receive simple prompts; the four Ollama models also receive complex prompts with definitions, IPIP phrases, and adjectives. Labels 0/1 are generated at temperature 0 and maximum 20 tokens. Responses of one digit or starting with 0/1 are accepted; all others are discarded. 135 classification reports are computed with accuracy, macro-F1, and per-class metrics. Pandora is rescaled and binarized with thresholds transferred from MyPersonality.

Sample: 2,467 Essays, 255 aggregated texts from MyPersonality, and 14,221 posts from Pandora are evaluated. MyPersonality classes sum to 250 per trait, not 255. Pandora contains multiple texts per 1,608 users and exhibits strong imbalance: except for Openness, positive classes are minority. Each model-dataset-trait-prompt combination is one experiment; there are 135, but invalid responses are excluded before measuring and dependence by user is not modeled.

Findings

  • The current source is arXiv:2511.23101v1, submitted on 28 November 2025, with 28 pages.
  • The 28 pages were rendered and visually inspected.
  • The current PDF matches byte for byte with the cache and has SHA-256 04dc74be5893c53bc5c743c22984bd9f9c01187d15bc8aed3645a50b03c08acd.
  • 75 experiments are conducted with the simple prompt and 60 with the complex prompt, 135 in total.
  • Only 13 configurations achieve F1 greater than or equal to 0.5 in both classes: 6 simple and 7 complex.
  • The pass rate for that cutoff is 9.6% total, 8.0% simple, and 11.7% complex.
  • The best balanced row is Mistral/Agreeableness/MyPersonality/complex with accuracy 0.608 and macro-F1 0.610.
  • That accuracy improves 0.072 over the majority class of the same sample.
  • GPT-4/Agreeableness/Essays obtains accuracy 0.593 and macro-F1 0.585 with its own prompt.
  • Mistral/Openness/Essays obtains accuracy 0.584 and macro-F1 0.575.
  • Openness and Agreeableness concentrate the majority of configurations that exceed the cutoff in both classes.
  • Extraversion and Neuroticism show the weakest and most biased results.
  • Complex prompts tend to raise recall of class 1 and reduce recall of class 0.
  • Complex prompts eliminate many invalid outputs, especially for Phi-3 on Essays.
  • Phi-3 produces between 80.18% and 84.84% invalid outputs on Essays with the simple prompt.
  • Llama 3.1 produces more invalid outputs on Pandora with the complex prompt than with the simple prompt.
  • Invalid outputs are discarded before accuracy and F1.
  • Table 2 reports 679 invalid outputs for GPT-4/Pandora for each trait.
  • Table 3 reports 480 unique invalid texts for GPT-4/Pandora and Table 4 also implies 480 for Openness.
  • The figures 679 and 480 are incompatible for the same condition.
  • MyPersonality declares 255 texts but each pair of classes sums to 250.
  • No code, raw predictions, processed Pandora labels, or execution logs were located.

Limitations

  • The Big Five are continuous dimensions; binary presence or absence simplifies and reifies the construct.
  • A high or low label from the author does not imply that each text expresses the trait.
  • The GPT-4 prompt defines 1 as a high or moderate level, distinct from presence or expression in the other prompts.
  • GPT-4 uses a different prompt and does not receive the complex condition.
  • Comparing complex open models with simple GPT-4 confounds architecture and prompt.
  • The GPT-4 snapshot and the date of the calls are not identified.
  • The estimate of one billion parameters for GPT-4 does not come from an official model specification.
  • The Ollama tags latest and instruct are mutable and do not guarantee the same artifact in a replication.
  • No hashes, quantization, chat templates, Ollama version, or model revisions are reported.
  • Temperature 0 does not guarantee determinism across runtimes, hardware, or backends.
  • No seed, top-p, stop tokens, or all generation parameters are reported.
  • Accepting any output that starts with 0 or 1 may classify an ambiguous or contradictory explanation as valid.
  • The parser is not published, so the exact treatment of spaces, signs, or subsequent text cannot be verified.
  • Discarding invalid outputs conditions the metrics on compliance and may favor models that abstain selectively.
  • No end-to-end performance is reported penalizing invalid outputs as errors or abstentions.
  • With 80–85% invalid outputs, the Phi-3/Essays simple results are based on a non-representative fraction.
  • The manual inspection of failures does not indicate subsample size, reviewers, criterion, or agreement.
  • Models discarded due to invalid formats are selected after a screening that is not fully documented.
  • The 0.5 cutoff is heuristic and not a significance test.
  • The paper acknowledges that 0.5 does not equal chance under imbalance.
  • There are no confidence intervals, bootstrap, paired tests, multiple correction, or effect size.
  • 135 configurations are inspected and the best ones are highlighted without controlling for multiple selection.
  • The best macro-F1 is 0.61 and does not by itself justify qualifying the system as reliable.
  • No ROC or precision-recall curves, calibration, Brier score, or probabilistic thresholds are compared.
  • The 0/1 predictions do not allow calibrating confidence or adjusting error costs.
  • Transferring thresholds from MyPersonality to Pandora assumes comparability between instruments and distributions that is not demonstrated.
  • The exact thresholds used for each Pandora trait are not published.
  • The phrase maximizes contrast is not formalized as a reproducible objective nor externally validated.
  • Pandora is strongly imbalanced due to the chosen binarization.
  • Pandora has 14,221 texts from 1,608 users; treating them as units may inflate apparent accuracy.
  • No bootstrap by user, clustered errors, or aggregation by author is used.
  • MyPersonality shows five fewer labels per trait than declared texts without explanation.
  • The MyPersonality invalid rates use 255 as the denominator even though the labeled supports are 250.
  • The inconsistency of 679 versus 480 prevents knowing the correct denominator for GPT-4/Openness/Pandora.
  • Table 1 does not clarify whether the central tendency of lengths is the mean or the median.
  • The lists in the complex prompt mix trait descriptors with morally loaded and socially desirable terms.
  • The higher positive recall may be a direct consequence of that polarity, not better psychological inference.
  • There is no separate ablation of definition, IPIP phrases, and adjectives.
  • No neutral, calibrated, counterbalanced, or alternative-order prompts are tested.
  • No sensitivity to wording, list order, or text length is measured.
  • The three datasets differ in pragmatic language, length, platform, and label process.
  • The cited SOTA benchmarks use different preprocessing, splits, supervised models, and metrics.
  • There is no evaluation on a common split that supports parity with previous systems.
  • No predictions, full confusion matrices for the 135 conditions, or scripts are published.
  • No license or reproducible access route for the exact version of MyPersonality used is discussed.
  • The task involves sensitive profiling; although the paper warns of risks, it does not evaluate demographic disparities or harm.
  • The study is a preprint with no evidence of peer review in the verified source.

What the study does not establish

  • It does not establish reliable personality prediction from text.
  • It does not demonstrate statistical significance with the F1 0.5 cutoff.
  • It does not demonstrate that 0 and 1 correspond to real absence and presence of a trait.
  • It does not validate the thresholds transferred from MyPersonality to Pandora.
  • It does not demonstrate that the complex prompt improves balanced performance in a general way.
  • It does not offer a controlled comparison between GPT-4 and the open models.
  • It does not demonstrate parity with supervised benchmarks under the same protocol.
  • It does not demonstrate independence of the Pandora observations.
  • It does not demonstrate that invalid output is random with respect to class or difficulty.
  • It does not allow resolving the inconsistency of 679 versus 480 without raw data.
  • It does not allow reproducing mutable models and unversioned GPT-4.
  • It does not justify use in employment selection, health, advertising, justice, or other sensitive contexts.

Traceability

Scope: Full text

Version: arXiv:2511.23101v1, submitted 28 November 2025, 28 pages

Consulted source: https://arxiv.org/abs/2511.23101

Review: Codex complete bilingual full-text fidelity pass, current arXiv-version reconciliation, all-page PDF visual inspection, model and dataset reproducibility audit, invalid-output denominator audit, class-wise table reconciliation, majority-baseline reconstruction, prompt-comparability and construct-validity assessment; summaries written from the full paper and reported tables rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4 through the OpenAI API; exact snapshot, run date and generation configuration beyond the prompt are not reported
  • phi3:latest via Ollama, described as approximately 3.8B; mutable tag without digest or quantization
  • gemma:7b-instruct via Ollama; mutable tag without digest or quantization
  • llama3.1:8b via Ollama; mutable tag without digest or quantization
  • mistral:instruct via Ollama, described as 7B; mutable tag without digest or quantization

Instruments and metrics

  • Binary Big Five/OCEAN labels for Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism
  • Simple open-model prompt asking whether the trait is expressed in the text
  • Different GPT-4 prompt mapping high or moderate trait level to 1 and low to 0
  • Complex prompt with Costa and McCrae high-level definitions, IPIP facet phrases and Goldberg adjective lists
  • Accuracy and macro-F1
  • Per-class precision, recall and F1
  • Invalid-output and shared-invalid-output counts
  • Heuristic F1 threshold of 0.5 for both classes

Data used

  • Essays: 2,467 stream-of-consciousness essays from 2,467 authors with near-balanced binary trait labels
  • MyPersonality: 255 reported Facebook users with status posts aggregated per author; only 250 class labels per trait are shown
  • Pandora: 14,221 Reddit texts from 1,608 users, with 1–100 scores rescaled and binarized against MyPersonality-derived thresholds
  • No released processed dataset snapshot, label file, model prediction file or parser output located

Evidence and location

  • Version, date, authors, and abstract: arXiv:2511.23101v1 metadata and title page checked 15 July 2026
  • Datasets, samples, distributions, and Pandora binarization: Section 3.1 and Table 1, pages 4–5
  • Models and exclusions: Section 3.2, page 5
  • Prompts, parameters, and response parser: Sections 3.3–3.4 and footnotes 11–13, pages 5–6; Appendix D, pages 21–28
  • Invalid outputs and cofailures: Tables 2–3 and Sections 4.1–4.2, pages 6–8
  • Discarding of invalid outputs and F1 0.5 cutoff: Section 4.3, page 8; Appendix A, page 14
  • Thirteen configurations and exact metrics: Figures 1–3, pages 9–10; Tables 4–5, pages 15–16
  • Effect of the complex prompt: Section 4.3.4 and Figures 4–8, pages 10 and 13, 17–20
  • Arithmetic reproduction of baselines and table contradictions: reports/verification/article-181-metric-and-design-audit.json
  • Absence of located artifacts: Paper, exact-title, arXiv-ID, GitHub and Zenodo searches checked 15 July 2026
  • Complete visual inspection: All 28 pages of arXiv:2511.23101v1 rendered and visually inspected on 15 July 2026