Can Large Language Models Assess Personality From Asynchronous Video Interviews? A Comprehensive Evaluation of Validity, Reliability, Fairness, and Rating Patterns

Evaluation and psychometric validity2024IEEE XploreApproved editorial review

Authors: Tianyi Zhang, Antonis Koutsoumpis, Janneke K. Oostrom, Djurre Holtrop, Sina Ghassemi, Reinout E. de Vries

Keywords: GPT-3.5, GPT-4, Asynchronous Video Interviews, HEXACO, Personality Assessment, Personnel Selection, Psychometric Validity, Reliability, Fairness, Reproducibility

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

6
Authors
16
Findings
26
Limitations
13
Evidence

Editorial summary

English

This study evaluates whether GPT-3.5 and GPT-4 can score personality and interview performance from responses elicited in asynchronous video interviews. Despite the video framing, the models receive only Google Speech-to-Text transcripts, the interview questions and, in the primary condition, information identifying the factor or facet targeted by each question. The final sample contains 685 participants answering eight questions for four Extraversion and four Conscientiousness facets in a fictional management-trainee application. Criteria are HEXACO-60 self-reports, ratings by four trained observers, and five interview ratings provided by two professional recruiters per participant. A ten-fold cross-validated supervised BoW-SBERT model serves as the comparator.

The positive result is concentrated in apparent Extraversion. Against observer ratings, GPT-4 reaches R²=.53 and GPT-3.5 .40, above BoW-SBERT's .36; GPT-4 also obtains R² values from .32 to .61 across the four Extraversion facets. The picture reverses for Conscientiousness: GPT-4 explains .11 and GPT-3.5 .08 of observer variance versus .23 for the baseline; both LLMs are below zero against self-report and also fail on Prudence and Perfectionism. Most importantly for hiring, every R² for the four recruiter-rated competencies and overall interview performance is negative, so the models perform worse than assigning the mean.

Reliability is measured only with Pearson correlation. When exactly the same text is scored twice, GPT-4 obtains r=.79 for Extraversion, .61 for Conscientiousness and .83 for interview performance; with a new interview 7-24 months later, the values are .59, .39 and .85. These correlations do not test absolute agreement, and the latter combines instrument inconsistency with genuine changes in responses. Raising temperature slightly increases observer-Extraversion R² while sharply reducing consistency. Scores are compressed and skew high. The fairness analysis compares gender differences and correlations with age, education and attractiveness, but it does not test measurement invariance, equal error, differential prediction or adverse impact; lower association than human observers is not sufficient evidence of fairness.

The candidate-selection simulation does not validate operational use. Its quantity labeled “accuracy” is actually recall or true-positive rate, and ties allow the selected set to exceed K; overlap therefore rises mechanically as K grows. There is no random-ranking baseline, specificity, decision utility or prospective workplace outcome. Semantic similarity between explanations and HEXACO items is also relative and partly circular because the prompt discloses the constructs and distances are normalized within each explanation.

The author preprint links a public repository containing 11,522 model-answer files and prediction artifacts at commit b2a3a70, a meaningful transparency contribution. It does not contain the OpenAI request code, exact model identifiers, complete messages, BoW-SBERT training, temperature/selection analyses or mixed-effects models. There is no README, license, dependency manifest, test suite or CI, and legacy pickles fail under pandas 2.2.3. The sole linguistic-analysis script has broken paths and variables and cannot produce the published figure. Its 685 output identifiers also link directly to a 710-row CSV containing demographics and psychometric ratings, without a reuse-governance note.

The faithful conclusion is narrow but useful: historical GPT-4 approximated apparent Extraversion well from text in this setting, but the study did not validate general personality assessment or employment decisions. Conscientiousness, job-performance validity, score agreement, fairness and operational utility remain weak or unestablished. The study itself supplies evidence against direct deployment: highly correlated interview scores can coexist with negative R².

Español

Este estudio evalúa si GPT-3.5 y GPT-4 pueden puntuar personalidad y desempeño de entrevista a partir de respuestas obtenidas en entrevistas de vídeo asíncronas. Aunque el título habla de vídeo, los modelos solo reciben transcripciones producidas por Google Speech-to-Text, las preguntas y, en la condición principal, información sobre el factor o la faceta que activa cada pregunta. La muestra final comprende 685 participantes que respondieron ocho preguntas para cuatro facetas de Extraversión y cuatro de Responsabilidad/Conscientiousness en una vacante ficticia de management trainee. Como criterios se usan el HEXACO-60 autorreportado, puntuaciones de cuatro observadores entrenados y cinco valoraciones de entrevista realizadas por dos reclutadores por persona. Un modelo supervisado BoW-SBERT entrenado con validación cruzada de diez pliegues sirve de comparación.

El resultado positivo se concentra en la Extraversión aparente. Frente a observadores, GPT-4 alcanza R²=0,53 y GPT-3.5 R²=0,40, por encima del 0,36 de BoW-SBERT; GPT-4 también obtiene R² entre 0,32 y 0,61 en las cuatro facetas de Extraversión. La imagen cambia para Conscientiousness: GPT-4 explica 0,11 y GPT-3.5 0,08 de la varianza observada, frente a 0,23 del baseline; ambos LLM quedan por debajo de cero contra el autoinforme y también fallan en Prudence y Perfectionism. Más importante para contratación, todos los R² de las cuatro competencias y de la puntuación global de entrevista son negativos: los modelos rinden peor que asignar la media.

La fiabilidad se mide únicamente con correlación de Pearson. Al repetir exactamente el mismo texto, GPT-4 obtiene r=0,79 en Extraversión, 0,61 en Conscientiousness y 0,83 en entrevista; con una nueva entrevista 7-24 meses después, baja a 0,59, 0,39 y 0,85. Estas correlaciones no prueban acuerdo absoluto y la última mezcla inestabilidad del instrumento con cambios reales en las respuestas. Subir la temperatura eleva ligeramente el R² de Extraversión y reduce con fuerza la consistencia. Las puntuaciones están comprimidas y sesgadas hacia valores altos. El análisis de fairness compara diferencias por género y correlaciones con edad, educación y atractivo, pero no estudia invariancia de medida, igualdad del error, predicción diferencial ni impacto adverso; una asociación menor que la de observadores no basta para demostrar equidad.

La simulación de selección tampoco valida su uso. La magnitud llamada «accuracy» es en realidad recall o tasa de verdaderos positivos, y los empates permiten seleccionar más de K personas; al aumentar K, el solapamiento sube mecánicamente. No hay baseline aleatorio, especificidad, utilidad ni resultados laborales posteriores. La similitud semántica entre explicaciones y ítems HEXACO es asimismo relativa y parcialmente circular: el prompt revela los constructos y normaliza las distancias dentro de cada explicación.

La prepublicación enlaza un repositorio público con 11.522 respuestas textuales y ficheros de predicción en el commit b2a3a70, una aportación de transparencia real. Sin embargo, no contiene el código de llamadas a OpenAI, identificadores exactos de modelo, mensajes completos, baseline BoW-SBERT, análisis de temperatura/selección ni modelos mixtos. No hay README, licencia, dependencias, tests o CI; los pickles antiguos fallan con pandas 2.2.3. El único script lingüístico tiene rutas y variables rotas y no puede producir la figura publicada. Además, sus 685 identificadores enlazan directamente con un CSV de 710 filas que incluye demografía y puntuaciones psicométricas, sin una nota de gobernanza de reutilización.

La conclusión fiel es limitada pero útil: el GPT-4 histórico aproximó bien la Extraversión aparente a partir de texto en este escenario, pero no validó una evaluación general de personalidad ni decisiones de empleo. Conscientiousness, validez para desempeño, acuerdo de las puntuaciones, fairness y utilidad operativa quedan débiles o sin establecer. El propio estudio aporta evidencia en contra del despliegue directo: alta consistencia de la entrevista puede coexistir con R² negativo.

Research question

With what validity, consistency, and demographic associations do GPT-3.5 and GPT-4 score Extraversion, Conscientiousness, and interview performance from AVI transcripts, how do they compare with observers and a supervised baseline, and how do they change with temperature and metainformation?

Method

Concurrent validation study on 685 simulated interviews for a fictitious vacancy. Eight questions activate four HEXACO facets of Extraversion and four of Conscientiousness. GPT-3.5 and GPT-4 receive only transcripts, score from 1 to 5 at temperature 1, and generate explanations. They are compared by R² with HEXACO-60 self-reports, four trained observers and two recruiters per participant, in addition to a BoW-SBERT MLP in ten-fold cross-validation. Consistency is estimated by correlating two runs of the same text and a new AVI of 145 persons after 7-24 months. Gender, age, education, and attractiveness, temperature 0-2, presence of metainformation, top-K overlap, and semantic similarity of explanations with HEXACO-60 are analyzed.

Sample: 685 final participants out of 947 recruited: 231 men, 447 women, and 7 non-binary persons; mean age 31.08 years (SD 11.52). 100 were excluded for corruption, 28 for attention, 92 for extreme variability in the inventory, and 42 manually flagged as non-compliant. A subgroup of 145 repeated the AVI 7-24 months later (mean 11.8; median 7).

Findings

  • Against observers, R² of GPT-3.5/GPT-4/BoW-SBERT is 0.40/0.53/0.36 for Extraversion and 0.08/0.11/0.23 for Conscientiousness.
  • Against self-report, Extraversion reaches 0.10/0.18/0.09; both LLMs have negative R² in Conscientiousness and the baseline reaches 0.03.
  • GPT-4 explains R²=0.61 of Sociability, 0.58 of Social Boldness, 0.44 of Social Self-Esteem, and 0.32 of Liveliness; it fails in Prudence and Perfectionism.
  • All R² for the four competencies and the global interview score are negative.
  • When repeating the same text, GPT-4 obtains r=0.79/0.61/0.83 in Extraversion, Conscientiousness, and interview; GPT-3.5, 0.68/0.48/0.75.
  • With a new AVI 7-24 months later, GPT-4 drops to 0.59/0.39 in personality and maintains 0.85 in interview despite this lacking concurrent validity; GPT-3.5 gives 0.48/0.19/0.78.
  • The standard deviation of GPT-4 is 0.55 in Extraversion and 0.41 in Conscientiousness, compared to 0.83/0.42 of observers and 0.68/0.54 of self-report; range compression exists.
  • When raising temperature from 0 to 2, the R² of Extraversion of GPT-4 increases approximately from 0.48 to 0.57, while the correlation drops from near 0.68 to 0.12.
  • Gender differences of GPT-3.5 are d=-0.374/-0.308 and those of GPT-4 -0.312/-0.395 for Extraversion/Conscientiousness; these effects alone do not identify measurement bias.
  • Attractiveness correlations of GPT-4 are 0.118/0.068 compared to 0.246/0.154 of observers, but a lower association does not prove equality of error or validity between groups.
  • The RECselect metric labeled as accuracy is recall; ties expand the selected set and the curve does not measure global classification accuracy.
  • The textual claim of less than 10% for K<10 is in tension with a recall point of GPT-4 without vacancy above 60% at K=5.
  • The article inverts the interpretation of the Shapiro-Wilk test: it calls non-normal results with p>0.05 and normal those with p<0.05.
  • The repository releases 11,522 outputs, but three GPT-3.5 folders are incomplete (682, 673, and 682 files) without the article publishing those n per condition.
  • The repository does not allow regenerating calls or figures and the linguistic script is not executable as versioned.
  • The 685 output identifiers match workerId in the CSV of 710 rows, which allows linking derived profiles with demographics and human measures.

Limitations

  • The LLMs only process transcribed text; the video or audio modality suggested by the title is not evaluated.
  • Only Extraversion and Conscientiousness activated by transparent questions are studied; the full HEXACO is not validated.
  • The prompts reveal factors or facets and ask to reason with HEXACO, so part of the semantic alignment is induced.
  • The exact snapshots of GPT-3.5 and GPT-4, execution dates, payloads, retries, or failure policy are not identified.
  • Pearson r measures relative order, not absolute agreement, individual error, or interchangeability of scores.
  • The retest interval varies between 7 and 24 months and mixes human change, new responses, and model instability.
  • Recruiter criteria have modest agreement, mean ICC 0.56, and are concurrent impressions, not future job performance.
  • All job R² are negative, so the top-K simulation lacks a predictive validity basis.
  • RECselect is recall although it is called accuracy; specificity, FPR, utility, calibration, or random baseline are not reported.
  • Ties make L≥K and can inflate recall by selecting many candidates, at the cost of precision.
  • The fairness analysis confounds score differences with bias and does not evaluate invariance, error by group, differential prediction, or adverse impact.
  • Observers, self-reports, and recruiters are not bias-free ground truths and measure partially different constructs.
  • The published interpretation of Shapiro-Wilk uses the wrong direction of p.
  • 92 participants are excluded for low/high dispersion in the inventory, a selection on the response pattern without reported sensitivity.
  • Semantic normalization forces a relative 0-1 scale within each explanation; it does not offer comparable absolute similarity.
  • Uncertainty intervals or sufficient stochastic repetitions for the temperature sweep are not shown.
  • The cost comparison omits transcription, prompt development, ground truth, validation, reruns, governance, and cost of error.
  • The repository has no README, license, requirements/lockfile, tests, CI, tags, or releases.
  • The generation code, BoW-SBERT baseline, temperature, selection, or mixed models are not released.
  • bias_analysis.py calculates descriptive effects and correlations, but not the mixed model described in the article.
  • linguistic_analysis.py overwrites the model variable, uses a nonexistent path, and divides arrays never populated; it does not reproduce Figure 8/Table III.
  • Inherited pickles depend on an old, undocumented version of pandas and fail under pandas 2.2.3.
  • Three GPT-3.5 conditions have fewer than 685 outputs and the management of missingness is not documented in the article.
  • The underlying OSF is not registered, is mutable, and corresponds mainly to the previous supervised study.
  • The IDs allow row-to-row linking of psychological outputs with demographics and ratings without a visible note of reuse, consent, or governance.
  • A fictitious vacancy, English, and a selected sample do not guarantee generalization to real jobs, languages, cultures, or populations.

What the study does not establish

  • It does not demonstrate that GPT-3.5 or GPT-4 validly assess personality in general.
  • It does not demonstrate prediction of job performance, hiring, retention, or future behavior.
  • It does not demonstrate that the model analyzes video, voice, facial expression, or audiovisual signals.
  • It does not demonstrate absolute agreement or individual stability by reporting high correlations.
  • It does not demonstrate fairness because a demographic or attractiveness correlation is lower than the human one.
  • It does not demonstrate that observers or recruiters are bias-free criteria.
  • It does not demonstrate that the explanations are faithful to the internal computation or that the model understands psychological theory.
  • It does not demonstrate that GPT-4 generally outperforms supervised models: it loses in Conscientiousness.
  • It does not validate selection utility; the so-called accuracy is recall under ties and all job R² are negative.
  • It does not allow regenerating the historical endpoints, prompts, and executions from the repository.
  • It does not allow end-to-end reproduction of all tables and figures with the released code.
  • It does not demonstrate that 50 euros is the total cost of a responsible evaluation system.
  • It does not automatically generalize outside the fictitious vacancy, English, the sample, or the 2023 versions.
  • It does not demonstrate that an LLM possesses personality, motives, self-knowledge, or stable psychological traits.

Traceability

Scope: Full text

Version: IEEE TAFFC 15(3), version of record dated 2 September 2024; university repository copy

Consulted source: https://research.vu.nl/ws/files/382311438/Can_Large_Language_Models_Assess_Personality_From_Asynchronous_Video_Interviews_A_Comprehensive_Evaluation_of_Validity_Reliability_Fairness_and_Rating_Patterns.pdf

Review: Codex full-text, visual, code, data-linkage and artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5 (exact API snapshot not reported)
  • GPT-4 (exact API snapshot not reported)
  • BoW-SBERT supervised regression baseline
  • SentenceTransformer all-MiniLM-L6-v2 (artifact linguistic script)

Instruments and metrics

  • HEXACO-60 self-report
  • HEXACO Behavioral Anchored Rating Scales (HEXACO-BARS)
  • Four job-related competencies and overall interview performance ratings
  • First-impression attractiveness rating
  • Coefficient of determination (R2)
  • Pearson repeated-score and long-interval retest correlation
  • Gender Cohen's d and demographic/attractiveness correlations

Data used

  • 685-person simulated asynchronous video interview dataset
  • 145-person repeat-interview subset
  • GitHub Tianyi-Zhang-TZ/LLMs_Personality at commit b2a3a703d15d6316ab3139f8fd0298fde96a23d3
  • OSF gsj46 underlying Prolific dataset and prior supervised-study materials

Evidence and location

  • Design, sample, exclusions, questions, and human criteria: Version of record, pp. 1772-1773, Methodology section III-A
  • Prompts, text-only modality, temperature, and baseline: Version of record, p. 1773, Figure 1, Figure 2 and sections III-B-C
  • R² by factor/facet, means, variances, and negative job results: Version of record, pp. 1775-1776, Figure 3 and section IV-A
  • Repetition and retest correlations: Version of record, p. 1776, Table I and section IV-B
  • Gender, age, attractiveness, and education: Version of record, pp. 1777-1778, Table II and section IV-C
  • Temperature and candidate selection: Version of record, pp. 1778-1779, Figures 4-6 and sections IV-D-E
  • Metainformation and linguistic similarity: Version of record, pp. 1780-1782, Figures 7-8, Table III and sections IV-F-G
  • Limits and caution of the authors themselves: Version of record, pp. 1782-1783, Discussion, Limitations and Conclusion
  • Link to the specific repository omitted in the VOR: Author preprint SHA-256 06ca97e7d7f8b0a9602a2cdff2484e41fa7c7c31b60ddef912cf0780539f833d, p. 2 footnote
  • Outputs, predictions, coverage, and absence of pipeline: GitHub https://github.com/Tianyi-Zhang-TZ/LLMs_Personality at commit b2a3a703d15d6316ab3139f8fd0298fde96a23d3, audited 16 Jul 2026
  • Errors of the linguistic script and absence of mixed models: GitHub commit b2a3a703d15d6316ab3139f8fd0298fde96a23d3: linguistic_analysis.py lines 11-114 and bias_analysis.py lines 16-124
  • Data, code, and supplement of the previous supervised study: OSF osf.io/gsj46 audited 16 Jul 2026; unregistered project, Supplemental Material.docx and codes.zip
  • Complete methodological and artifact report: reports/verification/article-210-avi-personality-validity-fairness-and-artifact-audit.json