Comparing chatbots to psychometric tests in hiring: reduced social desirability bias, but lower predictive validity

Applications, bias, and safety2025Frontiers (Psychology)Approved editorial review

Authors: Danilo Dukanovic, Dario Krpan

Keywords: AI personality assessment, Hiring, Chatbots, Big Five, Social desirability bias, Psychometric validity, Predictive validity, Propensity score matching, Algorithmic hiring, Reproducibility

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

2
Authors
20
Findings
45
Limitations
13
Evidence

Editorial summary

English

Comparing chatbots to psychometric tests in hiring compares a conversational Big Five assessment with a traditional questionnaire and asks whether the chatbot offers psychometric validity, reduced susceptibility to social desirability, and association with professional outcomes. It does not study a personality inside the LLM: a commercial OpenAI-based system scores human responses. The design is cross-sectional and quasi-experimental. Of 264 people entering the survey, 159 also completed the chatbot: 114 controls and 45 candidates. All 45 candidates came from one bank, two one-hour sessions and two roles; controls came from Recrewty clients or prospects and LinkedIn across multiple sectors. Before matching, groups differed in age, education, job level and industry. The authors explore several propensity-score methods and select a 0.25 caliper retaining 33 pairs, N=66. The system was built on Fabrile with Ingram, Recrewty and the authors. The article calls it a 'ChatGPT 4.0 API', says it was fine-tuned with existing data, and the supplement reveals that it had prior trait/facet means and standard deviations. It does not identify the exact API model, snapshot, system/scoring prompt, scoring rubric, fine-tuning dataset, training/evaluation separation, temperature, seed, retry policy or pipeline. It asks at least 15 questions, one per facet, with optional follow-ups. The published questions are transparent and leading: they ask when a person helped someone, completed a task, behaved responsibly, felt anxious or solved something creatively. They therefore do not demonstrate covert extraction of authentic patterns or resistance to gaming. Benchmarks are a 50-item Serbian Big Five plus Two short form and a 13-item Marlowe-Crowne scale. Twenty-one missing item responses are mean-imputed. The questionnaire Big Five CFA, using only 159 cases for 50 items, yields CFI 0.705 and TLI 0.690, although the paper calls it acceptable by emphasizing RMSEA 0.074 and GFI 0.968. The chatbot CFA over 15 facets yields CFI 0.905, TLI 0.875, RMSEA 0.078 and GFI 0.880. Reliability is good for Extraversion (alpha 0.80) and Neuroticism (0.85), poor for Agreeableness (0.58), and modest for Conscientiousness (0.67) and Openness (0.64). Convergence with traditional tests is r=0.443 for Extraversion, 0.449 for Conscientiousness, 0.362 for Openness, 0.256 for Agreeableness and only 0.089, non-significant, for Neuroticism. Discriminant validity is problematic: AI Agreeableness correlates more with traditional Extraversion (0.342) and Neuroticism (-0.272) than with Agreeableness (0.256), with other cross-trait associations. Claimed substantive validity uses facet-total correlations where the total contains that facet and extreme-group discrimination defined by the same composite; these are circular part-whole checks, not independent content validation. AI scores provide no useful prediction or incremental value for education or job level. Education and job level are concurrent demographics, not future performance, hiring decisions, retention or supervisor ratings; calling them predictive validity or real-world outcomes overstates the criterion. The headline lower-social-desirability claim is not established. Within 33 matched pairs, some traditional-test outcomes differ between candidates and controls while all five AI regressions are non-significant. But significance in one method and non-significance in another does not show that effects differ. There is no treatment-by-method interaction, formal difference test, equivalence/non-inferiority test, power for accepting the null, or shared social-desirability indicator applied to both. Marlowe-Crowne is only measured as a human questionnaire; AI bias is inferred from Big Five group differences. AI effects are highly imprecise: estimate(SE) -0.133(4.643) for Openness, 0.007(1.976) for Conscientiousness, -2.405(2.702) for Extraversion, 1.817(1.754) for Agreeableness and 2.675(10.236) for Neuroticism. Approximate intervals allow meaningful effects in either direction. Selection context is not isolated: every candidate shares bank, roles, sessions, recruiter and source, unlike controls. AI O/C/E/A means are 73.7-80.5 with strong negative skew and ceilings near 90; such compression, potentially influenced by undocumented priors, can hide group differences. The audit also finds published contradictions that block reliable significance interpretation. Openness is printed as 0.448 with SE 0.597 yet marked p<0.05, which is arithmetically incompatible. HAC 0.640(SE 0.173) appears under Agreeableness while prose calls it Extraversion. The AI education table shows Openness 0.05138(SE 0.02374), ratio about 2.16, while prose says every effect is non-significant. The supplement prints impossible Shapiro-Wilk W=2.038; the job-level model reports treatment -19.720 with SE 1511; and a figure captioned Neuroticism has an internal Extraversion_AI title. Without data or code, the correct results cannot be recovered. No article-specific repository, dataset, scripts, prompts or reproducible release was found on the official page or in targeted public GitHub, OSF and Zenodo searches. The data statement only promises access on request and part of the system remains proprietary. The faithful conclusion is narrower: this small sample detected no candidate-control difference in five AI scores and found no predictive or incremental AI validity, with moderate convergence for selected dimensions and clear failures for others. It does not show reduced bias. Employment use would require a frozen inspectable model/prompt/scoring specification, preregistration, within-person manipulation or multi-site randomization, a common direct faking measure, interaction and equivalence tests, pair/site-aware inference, prospective performance outcomes, invariance/test-retest/subgroup/adverse-impact studies, and reproducible artifacts.

Español

Comparing chatbots to psychometric tests in hiring compara una evaluación conversacional de Big Five con un cuestionario tradicional y pregunta si el chatbot ofrece validez psicométrica, menor susceptibilidad a deseabilidad social y asociación con resultados profesionales. No estudia una personalidad interna del LLM: un sistema comercial basado en OpenAI puntúa respuestas de personas. El diseño es transversal y cuasiexperimental. De 264 personas que iniciaron la encuesta, 159 completaron también el chatbot: 114 controles y 45 candidatos. Los 45 candidatos proceden de un único banco, dos sesiones de una hora y dos puestos; los controles proceden de clientes o prospectos de Recrewty y LinkedIn, con sectores variados. Antes del matching, candidatos y controles difieren en edad, educación, nivel profesional e industria. Los autores prueban varios métodos de propensity-score matching y eligen caliper 0,25, que conserva 33 parejas, N=66. El sistema se construyó en Fabrile con Ingram, Recrewty y los autores. El artículo lo llama 'ChatGPT 4.0 API', dice que fue fine-tuned con datos existentes y el suplemento revela que disponía de medias y desviaciones de rasgos/facetas previas. No identifica modelo/API exactos, snapshot, system prompt, rúbrica de scoring, dataset de fine-tuning, separación train/evaluación, temperatura, seed, retries ni pipeline. Formula al menos 15 preguntas, una por faceta, con follow-up opcional. Las preguntas publicadas son transparentes y dirigidas: piden recordar cuándo la persona ayudó, terminó una tarea, fue responsable, sintió ansiedad o resolvió algo creativamente. Por eso no demuestran extracción encubierta de patrones auténticos ni resistencia a gaming. Como benchmarks se usan un Big Five plus Two serbio corto de 50 ítems y Marlowe-Crowne de 13 ítems. Hay 21 valores ausentes imputados con la media del ítem. El CFA del cuestionario Big Five, con solo 159 casos y 50 ítems, da CFI 0,705 y TLI 0,690, aunque el paper lo llama aceptable apoyándose en RMSEA 0,074 y GFI 0,968. El CFA del chatbot sobre 15 facetas da CFI 0,905, TLI 0,875, RMSEA 0,078 y GFI 0,880. La fiabilidad es buena para Extraversión (alpha 0,80) y Neuroticismo (0,85), baja para Amabilidad (0,58) y modesta para Responsabilidad (0,67) y Apertura (0,64). La convergencia con el test tradicional es r=0,443 para Extraversión, 0,449 para Responsabilidad, 0,362 para Apertura, 0,256 para Amabilidad y solo 0,089 no significativa para Neuroticismo. La validez discriminante falla: AI-Amabilidad correlaciona más con Extraversión tradicional (0,342) y Neuroticismo (-0,272) que con Amabilidad (0,256), y hay más cruces. La llamada validez sustantiva usa correlaciones faceta-total donde el total incluye la misma faceta y discriminación por grupos extremos definidos con ese composite; son pruebas circulares/part-whole, no validación independiente de contenido. Los scores AI no predicen ni añaden valor útil para educación o nivel de puesto. Además, educación y job level son datos concurrentes, no desempeño futuro, decisión de contratación, retención o evaluación de supervisor; llamarlos predictive validity o real-world outcomes exagera el criterio. El claim principal de menor deseabilidad social no queda demostrado. En las 33 parejas, algunos tests tradicionales cambian entre candidatos y controles y las cinco regresiones AI no son significativas. Pero significativo en un método y no significativo en otro no prueba que los efectos difieran. No hay interacción tratamiento×método, test formal de diferencia, equivalencia/no inferioridad, potencia para aceptar el null ni un indicador común de deseabilidad social aplicado a ambos. Marlowe-Crowne solo se mide como cuestionario humano; para AI se infiere bias a partir de diferencias en Big Five. Los efectos AI son muy imprecisos: estimación(SE) -0,133(4,643) Apertura, 0,007(1,976) Responsabilidad, -2,405(2,702) Extraversión, 1,817(1,754) Amabilidad y 2,675(10,236) Neuroticismo. Sus intervalos aproximados permiten efectos relevantes en ambas direcciones. Tampoco se aísla el contexto de selección: todos los candidatos comparten banco, roles, sesiones, recruiter y fuente, mientras los controles no. Los scores AI de O/C/E/A tienen medias 73,7-80,5, fuerte sesgo negativo y techo cercano a 90; esa compresión, quizá influida por priors no documentados, puede ocultar diferencias. La auditoría descubre contradicciones publicadas que bloquean una lectura de significación fiable. Openness aparece con 0,448 y SE 0,597 pero lleva asterisco p<0,05, aritméticamente incompatible. El HAC 0,640(SE 0,173) está bajo la columna Amabilidad aunque el texto lo llama Extraversión. La tabla de educación AI muestra Apertura 0,05138(SE 0,02374), ratio ~2,16, aunque el texto dice que todo es no significativo. El suplemento imprime Shapiro-Wilk W=2,038, imposible porque W<=1; el modelo de nivel laboral da treatment -19,720 con SE 1511; y la figura captioned Neuroticism contiene el título interno Extraversion_AI. Sin datos ni código no puede resolverse cuál resultado es correcto. No se encontró repositorio, dataset, scripts, prompts o release reproducible específico del artículo en la página oficial ni en búsquedas públicas dirigidas de GitHub, OSF y Zenodo. El data statement solo promete disponibilidad bajo petición y parte del sistema permanece propietaria. La conclusión fiel es más limitada: en esta muestra pequeña no se detectó una diferencia candidato-control en los cinco scores AI y el chatbot no mostró validez predictiva/incremental; sí hubo convergencia moderada para algunas dimensiones y fallos claros para otras. No se demuestra reducción de bias. Antes de uso laboral hacen falta modelo/prompt/scoring congelados e inspeccionables, preregistro, manipulación within-person o randomización multi-sede, medida común directa de faking, interacción y equivalence tests, inferencia por parejas/sedes, outcomes prospectivos de desempeño, invariance/test-retest/subgrupos/adverse impact y artefactos reproducibles.

Research question

Can an LLM-based chatbot infer Big Five facets with acceptable structure, content, and external validity, predict professional outcomes, and show lower social desirability distortion than a traditional questionnaire in a selection context?

Method

Quasiexperimental cross-sectional study of 159 persons who complete a 50-item Serbian Big Five, 13-item Marlowe-Crowne, and a Fabrile/OpenAI chatbot with 15 open-ended questions directed at facets. CFA, alpha, facet-total correlations, and AI-test are applied, along with demographic tests, multinomial/logistic regressions, and propensity-score matching. For H2, 33 pairs are selected with a caliper of 0.25, and traditional and AI scores are compared via separate regressions with covariates. The audit read and rendered the 17 pages of the VOR and the 21 pages of the official supplement, verifying questions, tables, arithmetic, construct scope, confounding, reproducibility, and public availability of code/data.

Sample: 264 persons started the survey and 159 completed the chatbot, so 105 (39.8%) are excluded without attrition analysis: 114 controls and 45 candidates. All candidates are from one bank and two sessions for Director of Exposition and Affluent Associate. The final matching retains 33 pairs, N=66. The initial groups differ in age, education, job level, and industry. There are 21 item responses imputed with the mean and seven ages reconstructed from year of birth.

Findings

  • The traditional Big Five CFA reports CFI 0.705 and TLI 0.690, very low despite the paper's favorable interpretation.
  • The chatbot CFA reports CFI 0.905, TLI 0.875, RMSEA 0.078, and GFI 0.880.
  • Alpha is 0.80 for Extraversion and 0.85 for Neuroticism, but 0.58 for Agreeableness, 0.67 for Conscientiousness, and 0.64 for Openness.
  • AI-test convergence is moderate for E 0.443, C 0.449, and O 0.362, weak for A 0.256, and absent for N 0.089.
  • AI-Agreeableness correlates more with Extraversion 0.342 and Neuroticism -0.272 than with Agreeableness 0.256.
  • The facet-total correlations of substantive validity contain the facet in the total and are inflated by part-whole.
  • AI scores do not provide useful incremental validity for education or professional level.
  • Education and job level are concurrent criteria, not future outcomes of performance or hiring.
  • In the 33 pairs, no significant differences are detected in the five AI scores.
  • It is not proven that treatment effects differ between AI and traditional tests.
  • No equivalence testing is performed, nor is social desirability measured directly in the chatbot.
  • The implicit intervals of the AI effects are wide and allow for practically relevant differences.
  • All candidates from a single bank leave selection confounded with organization, roles, sessions, and source.
  • The chatbot questions are explicit and socially transparent, so they can be gamed.
  • AI scores show high means and ceiling, a possible alternative explanation for the null between groups.
  • Several tables contradict text, asterisks, or trait identity; one W statistic is impossible.
  • No exact model, prompts/scoring, fine-tuning data, parameters, code, or raw data are published.
  • No article-specific public reproducible artifact was found as of 16 July 2026.
  • The negative result on predictive/incremental validity is informative and should be retained without converting the null of bias into positive evidence.
  • The study evaluates AI-mediated scores from humans, not the internal personality of the LLM.

Limitations

  • Cross-sectional and non-randomized design.
  • Only 45 candidates and 33 pairs in H2.
  • All candidates belong to one bank, two sessions, and two roles.
  • Control recruited from different sources and sectors.
  • Matching does not eliminate confounding by bank, site, recruiter, session, or role.
  • The caliper is chosen after exploring alternatives and retaining the recommended N.
  • There is no preregistration or separate confirmatory analysis.
  • A difference between methods is inferred from separate significance patterns.
  • There is no treatment-by-method interaction or direct test of coefficients.
  • Equivalence/non-inferiority testing for the AI nulls is missing.
  • There is no Marlowe-Crowne or common faking measure applied to AI output.
  • The AI treatment intervals are very wide.
  • Ceiling and skew in AI scores may hide differences.
  • Mean/SD priors enter the chatbot in an undocumented manner.
  • Questions reveal the facet and assume desirable behaviors.
  • There is no exact model or API snapshot.
  • There is no system/scoring prompt, rubric, or sampling parameters.
  • Fine-tuning dataset and train/test separation are not described.
  • Stochastic stability, test-retest, or version change is not evaluated.
  • Traditional CFA has very poor CFI/TLI.
  • Chatbot CFA has TLI/GFI below 0.90.
  • Low/modest reliability in three of five AI traits.
  • Deficient discriminant validity.
  • Neuroticism lacks convergence.
  • Facet-total and extreme groups are circular.
  • Education/job level are not future job performance.
  • There are no hiring, supervisor, productivity, retention, or adverse impact outcomes.
  • 105 of 264 do not complete the chatbot and there is no attrition analysis.
  • 21 missing values are imputed with means without sensitivity analysis.
  • N=66 regressions include multiple covariates and may be overfitted.
  • No cluster/pair-aware inference is reported after matching.
  • Durbin-Watson/HAC on cross-sectional rows does not address the natural dependence of pairs/site.
  • There is no multiplicity correction.
  • The Openness asterisk of 0.448/SE 0.597 is incompatible with p<0.05.
  • HAC 0.640 appears in column A but the text attributes it to E.
  • AI Openness education 0.05138/SE 0.02374 contradicts the all non-significant claim.
  • Shapiro-Wilk W=2.038 is impossible.
  • Job-level treatment SE=1511 indicates an unstable/separated model.
  • The Neuroticism figure carries the internal title Extraversion_AI.
  • Raw data, transcripts, code, prompts, and release are not linked.
  • The data statement is future availability upon request, not a persistent release.
  • The first authorship has a declared commercial conflict of interest with Recrewty.
  • Part of the method/product remains proprietary.
  • There is no independent replication, multi-site, or external validation.
  • Fairness by gender, age, education, language, or other subgroups is not studied.

What the study does not establish

  • It does not demonstrate that the chatbot reduces social desirability or faking.
  • It does not demonstrate that the selection effect is smaller in AI than in questionnaires.
  • It does not demonstrate a statistical difference between both methods.
  • It does not demonstrate equivalence between candidates and controls.
  • It does not demonstrate causality of the selection context.
  • It does not separate bank, position, session, recruiter, and source from the treatment.
  • It does not demonstrate predictive validity for future job performance.
  • It does not demonstrate acceptable validity and reliability in all five traits.
  • It does not demonstrate resistance to gaming of transparent questions.
  • It does not provide independent content validation.
  • It does not allow reproducing the chatbot scoring.
  • It does not allow deciding which contradictory version of the results is correct.
  • It does not demonstrate public availability of data/code.
  • It does not generalize beyond the bank, roles, language, region, and system evaluated.
  • It does not demonstrate fairness or absence of adverse impact.
  • It does not demonstrate personality, stable trait, or mental state within the LLM.

Traceability

Scope: Full text

Version: Frontiers in Psychology version of record, volume 16 article 1564979, published 25 April 2025; audited with the official supplementary DOCX rendered to 21 pages; public code, data and model-artifact searches completed 16 July 2026

Consulted source: https://doi.org/10.3389/fpsyg.2025.1564979

Review: Codex complete bilingual fidelity pass using the full 17-page Frontiers version of record and official supplementary DOCX rendered to 21 pages, complete all-page visual inspection, DOI and metadata verification, question/transcript/table/arithmetic/construct audit, causal and statistical review of the social-desirability claim, and targeted public code/data/model-artifact searches; summaries written from full evidence rather than abstract keywords, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Unspecified OpenAI model described as ChatGPT 4.0 API
  • Custom Fabrile chatbot developed with Ingram Technologies and Recrewty
  • Five-factor confirmatory factor analysis with MLR
  • Multinomial logistic regression
  • Binary logistic regression
  • Ordinary linear regression
  • HAC standard-error regression
  • Robust regression using robustbase
  • Propensity-score matching with nearest-neighbor, caliper, full and optimal variants

Instruments and metrics

  • 50-item Serbian Big Five plus Two short form
  • 13-item short Marlowe-Crowne Social Desirability Scale
  • 15 chatbot Big Five facet questions with optional follow-ups
  • Confirmatory factor analysis
  • Cronbach alpha
  • Facet-total correlation
  • Extreme-group discrimination index
  • Convergent and discriminant correlation matrix
  • Education and dichotomized job level as concurrent criteria
  • Propensity-score matching
  • Shapiro-Wilk, Mardia, Durbin-Watson and VIF diagnostics

Data used

  • Study survey and chatbot responses: 159 completers from Serbia and Montenegro; not publicly linked
  • Professional-selection group: 45 candidates from one bank, two sessions and two roles
  • Control group: 114 professionals recruited through Recrewty contacts and LinkedIn
  • Matched analysis subset: 33 treated-control pairs, N=66
  • Prior Big Five facet means and standard deviations from Smederevac/Colovic-related work supplied to the chatbot; exact scoring use not documented
  • Official supplementary DOCX with questions, one transcript, tables and figures

Evidence and location

  • Metadata, abstract, question, and main claims: Frontiers in Psychology version of record pages 1-4; DOI 10.3389/fpsyg.2025.1564979
  • Chatbot, facet questions, ambiguous model, and fine-tuning: Version of record pages 5-6; official supplement rendered pages 4-13
  • Sample, one bank, two sessions, and attrition: Version of record pages 6-8; Table 1 and Procedure
  • Traditional/chatbot CFA and reliability: Version of record pages 8-10; Tables 2, 4 and 5
  • Convergence, discrimination, and external validity: Version of record pages 9-11; Tables 6-8
  • Propensity matching and balance: Version of record pages 11-12; Tables 9-11
  • Traditional regressions, HAC, and trait contradiction: Version of record pages 12-13, Table 12; official supplement rendered page 17, Table 7
  • AI nulls and standard error width: Version of record page 13; official supplement rendered pages 17-18, Table 8
  • Table errors, impossible W, unstable model, and mislabeled figure: Official supplement rendered pages 14-21, Tables 2-8 and Figures 1-5
  • Limitations, data, conflict, and proprietary material: Version of record pages 15-16, Limitations, Data availability and Conflict of interest
  • Public reproducibility search: Official article links plus targeted GitHub, OSF and Zenodo searches completed 16 July 2026; no article-specific reproducible release found
  • Comprehensive validity and reproducibility report: reports/verification/article-208-chatbot-hiring-bias-validity-and-reproducibility-audit.json
  • Complete visual inspection: All 17 version-of-record pages and all 21 pages rendered from the official supplementary DOCX visually inspected on 16 July 2026