Automatic Item Generation for Personality Situational Judgment Tests with Large Language Models

Evaluation and psychometric validity2026ElsevierApproved editorial review

Authors: Chang-Jin Li, Jiyuan Zhang, Yun Tang, Jian Li

Keywords: Large Language Models, Personality Assessment, Situational Judgment Tests, Psychometrics, Big Five, Automatic Item Generation, Content Validity, Test-Retest Reliability, Prompt Engineering, Response Position Bias

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

4
Authors
16
Findings
23
Limitations
12
Evidence

Editorial summary

English

Automatic Item Generation for Personality Situational Judgment Tests with Large Language Models asks whether an LLM can assist in building situational judgment tests (SJTs) that measure personality in people. It does not measure the model's personality, identity, or consciousness; its subject is automated generation of Chinese psychometric items. The program contains three studies. In Study 1, seven psychology doctoral students rate 21 translated human items and seven-item groups generated under four temperatures, three prompt variants, and two time points. Outcomes are situation necessity through CVR/CVI, option rationality, scoring rationality, and overall quality. Temperature 1.0 and 1.1 tie at CVI 0.714, with 1.0 obtaining stronger ranks for necessity and options. Prompt v2 reaches the highest CVI, 0.755, but Prompt v1 has higher mean overall quality, 6.57 versus 6.00, and v1-to-v2 gains are not significant. Selecting v2/1.0 is therefore an exploratory decision based on seven items per condition and the same seven judges, not confirmed optimization in an independent sample. The temporal comparison also does not establish stability: failure to detect differences between two seven-item groups generated ten days apart is neither an equivalence test nor regeneration of the same items. Study 2 generates 200 items through the consumer ChatGPT instant interface dated 6 November 2025 and labelled ChatGPT-5: five facets, five conversations/rounds, and eight items per cell. Overall CVI is 0.707; facet CVIs are 0.707 for self-consciousness, 0.657 gregariousness, 0.843 openness to ideas, 0.600 compliance, and 0.729 self-discipline. No mean round effect is detected, but this is not equivalent to reproducibility. Fifty of 200 items fall below CVR 0.71, 59 score below 6/7 overall quality, and 32 have mean scoring rationality below 3.5/4. Some failures are substantive: A_1_5 has CVR -1, scoring rationality 1.86, and quality 0/7; E_4_5 also has CVR -1 and quality 0/7. The paper appropriately concludes that expert screening remains necessary. Its cross-model claim is weaker: seven GPT-4 items from one facet are descriptively compared with 200 ChatGPT-5 items from five facets, different dates, and different interfaces, without a matched model contrast. The supplement also reveals that Studies 1 and 3 did not call OpenAI directly. They used `ai.shanxl.com`, a third-party gateway claiming to expose `gpt-4-1106-preview`. No code, provider receipt, request IDs, or logs verify the backend or effective parameters. Study 2 likewise lacks an immutable ChatGPT-5 backend ID, system prompt, seed, or conversation export. Study 3 creates a separate 40-item GPT-4 form, eight items for one selected facet from each Big Five domain, and administers it with corresponding NEO-PI-R facets. There are 443 valid cases, 130 with criteria, and 80 with a two-week retest. Open data reproduce the results: alpha 0.75/0.84/0.70/0.57/0.61, retest 0.58/0.77/0.41/0.52/0.65, and matched NEO convergence 0.68/0.67/0.48/0.36/0.44 for N/E/O/A/C. Evidence is therefore reasonable for self-consciousness, gregariousness, and partly openness, but compliance and self-discipline have low consistency, openness retest is 0.41, and compliance converges at only 0.36. WLSMV CFA reports RMSEA 0.03, CFI/TLI 0.95, and SRMR 0.11; the last is weak. Comparing mean inter-factor correlations, 0.385 for the LLM-SJT versus 0.494 for NEO, does not by itself establish discriminant validity: LLM compliance correlates 0.39 with NEO gregariousness and openness, above its 0.36 matched compliance correlation. Of 35 LLM-SJT criterion correlations, eight have unadjusted p<0.05 and only four survive a simple Bonferroni threshold; the paper does not pre-specify a multiplicity correction. The OSF release is a meaningful contribution: CC BY 4.0, the complete 310-item bank, prompts, supplement, `.sav` data, `.spv` outputs, and CFA inputs. All five datasets have no missing cells, IDs are unique, the 130 criterion and 80 retest IDs link to the 443 cases, and the audit reproduces CVR, means, sums, alphas, correlations, and ANOVA. No exact duplicate stems or option sets were found. Yet the manuscript says analysis scripts are public while OSF actually provides `.spv` viewers rather than SPSS syntax, plus two Mplus `.inp` files without `.out`; the exact pipeline is not executable end to end. Generation, timing, token, and cost logs are also absent. The claim of producing 100 items for $0.02-$1.00 and within minutes is an estimate, not a measurement of the gateway workflow used. Every generated item puts high-trait responses in A/B and low-trait responses in C/D, creating a transparent positional key that is not tested against coaching, faking, or social desirability. The work covers five selected facets rather than the complete Big Five, and cultural appropriateness is limited to Chinese judges and participants without cross-cultural invariance. The faithful conclusion is that the study offers valuable feasibility evidence, open data with good internal integrity, and partial psychometric success. It does not establish a universal, reproducible, or uniformly valid generator, and it does not establish synthetic personality. Verifiable model provenance, executable scripts, logs, equivalence testing, randomized response positions, independent replication, expert screening, and broader cultural validation are still needed.

Español

Automatic Item Generation for Personality Situational Judgment Tests with Large Language Models estudia si un LLM puede ayudar a construir pruebas de juicio situacional (SJT) para medir personalidad en personas. No mide personalidad, identidad ni consciencia del modelo: su objeto es la generación automática de ítems psicométricos en chino. El programa contiene tres estudios. En el primero, siete doctorandos en psicología valoran 21 ítems humanos traducidos y grupos de siete ítems generados bajo cuatro temperaturas, tres variantes de prompt y dos momentos. Los indicadores son necesidad de la situación mediante CVR/CVI, racionalidad de opciones, racionalidad de puntuación y calidad global. Temperatura 1,0 y 1,1 empatan en CVI, 0,714; 1,0 obtiene mejores rangos en necesidad y opciones. Prompt v2 logra el CVI más alto, 0,755, pero Prompt v1 tiene mejor calidad global media, 6,57 frente a 6,00, y las mejoras v1→v2 no son significativas. La selección de v2/1,0 es por tanto una decisión exploratoria sobre siete ítems por condición y los mismos siete jueces, no una optimización confirmada en una muestra independiente. La comparación temporal tampoco demuestra estabilidad: que dos grupos de siete ítems generados con diez días de separación no difieran significativamente no es una prueba de equivalencia ni de reproducción de los mismos ítems. El segundo estudio genera 200 ítems con la interfaz web de ChatGPT en modo instant del 6 de noviembre de 2025, denominada ChatGPT-5: cinco facetas, cinco conversaciones/rondas y ocho ítems por celda. Su CVI global es 0,707; por faceta es 0,707 en autoconciencia, 0,657 en gregarismo, 0,843 en apertura a ideas, 0,600 en complacencia y 0,729 en autodisciplina. No se detecta efecto medio de ronda, pero eso no equivale a reproducibilidad. Cincuenta de 200 ítems quedan bajo el umbral CVR 0,71 y 59 reciben menos de 6/7 en calidad. Treinta y dos tienen racionalidad de puntuación media menor de 3,5/4. Algunos fallos son sustantivos: A_1_5 obtiene CVR -1, puntuación 1,86 y calidad 0/7; E_4_5 también CVR -1 y calidad 0/7. El propio paper concluye correctamente que el filtrado experto sigue siendo necesario. La afirmación de generalización entre modelos es más débil: compara descriptivamente siete ítems GPT-4 de una sola faceta con 200 ítems ChatGPT-5 de cinco facetas, fechas e interfaces distintas, sin un contraste emparejado de modelo. Además, el suplemento revela que los estudios 1 y 3 no llamaron directamente a OpenAI: usaron `ai.shanxl.com`, una pasarela de terceros que afirmaba exponer `gpt-4-1106-preview`. No hay código, recibos, request IDs ni logs que verifiquen el backend o sus parámetros. En Study 2 tampoco se publica un identificador inmutable del backend de ChatGPT-5, system prompt, seed o exportación de conversaciones. El tercer estudio crea otros 40 ítems GPT-4, ocho por una única faceta de cada dominio Big Five, y los administra junto con facetas NEO-PI-R. Hay 443 casos válidos, 130 con criterios y 80 con retest de dos semanas. Los datos abiertos reproducen los resultados: alpha 0,75/0,84/0,70/0,57/0,61, retest 0,58/0,77/0,41/0,52/0,65 y convergencia con NEO 0,68/0,67/0,48/0,36/0,44 para N/E/O/A/C. Por tanto hay evidencia razonable para autoconciencia, gregarismo y en parte apertura, pero complacencia y autodisciplina tienen consistencia baja, apertura tiene retest 0,41 y complacencia converge solo 0,36. El CFA WLSMV informa RMSEA 0,03, CFI/TLI 0,95 y SRMR 0,11; este último es débil. La comparación de correlaciones interfactoriales medias, 0,385 en LLM-SJT frente a 0,494 en NEO, no establece por sí sola validez discriminante: la complacencia LLM correlaciona 0,39 con gregarismo y apertura NEO, más que 0,36 con complacencia NEO. De 35 correlaciones LLM-SJT con criterios, ocho tienen p<0,05 sin corregir y solo cuatro sobreviven un Bonferroni simple; el paper no preespecifica ajuste por multiplicidad. Los archivos OSF constituyen una aportación importante: licencia CC-BY-4.0, banco completo de 310 ítems, prompts, suplemento, datos `.sav`, salidas `.spv` e inputs CFA. Las cinco bases no tienen valores ausentes, los IDs son únicos, los 130 criterios y 80 retests enlazan con los 443 casos, y la auditoría reproduce CVR, medias, sumas, alphas, correlaciones y ANOVA. No hay duplicados exactos de stems u opciones. Pero el manuscrito dice publicar analysis scripts y en realidad ofrece viewers `.spv`, no sintaxis SPSS, además de dos `.inp` Mplus sin `.out`; la pipeline exacta no es ejecutable de extremo a extremo. Tampoco hay logs de generación, tiempos, tokens o costes. La afirmación de generar 100 ítems por 0,02-1,00 dólares y en pocos minutos es una estimación, no una medición de la pasarela usada. Todos los ítems generados colocan las respuestas de rasgo alto en A/B y bajo en C/D, una clave posicional transparente no evaluada frente a coaching, faking o deseabilidad social. El trabajo cubre cinco facetas, no todas las dimensiones/facetas Big Five, y la adecuación cultural se limita a jueces y participantes chinos, sin invariancia intercultural. La conclusión fiel es que el estudio ofrece evidencia de factibilidad valiosa, datos abiertos de buena integridad y éxito psicométrico parcial. No prueba un generador universal, reproducible o uniformemente válido, ni personalidad sintética; requiere procedencia verificable del modelo, scripts ejecutables, logs, tests de equivalencia, posiciones de respuesta aleatorizadas, replicación independiente, cribado experto y validación cultural más amplia.

Research question

Can a structured prompting procedure automatically generate Chinese SJT items for five personality facets with content validity, mean quality across rounds, and sufficient psychometric properties for use as a human instrument?

Method

Study 1 compares 21 translated human SJT items with groups of seven items generated through a third-party gateway labeled `gpt-4-1106-preview`, four temperatures, three prompts, and two time points; seven doctoral students score CVR/CVI, options, scoring, and quality. Study 2 uses the ChatGPT instant interface of 6-11-2025, called ChatGPT-5, for five independent rounds of 40 items (eight per five facets), and seven experts re-evaluate 200 items. Study 3 generates 40 new items with the chosen configuration, administers them with five NEO-PI-R facets to 443 people, adds seven criteria in 130 and retest in 80, and analyzes items, alpha, Pearson, CFA WLSMV, and correlations. The audit visually reviewed the 67 pages of the manuscript and the 16 of the supplement, downloaded all OSF, recounted banks and samples, verified links, reproduced indicators and ANOVA, inspected prompts/CFA/outputs, and evaluated model provenance, multiplicity, equivalence, faking, culture, and reproducibility.

Sample: Study 1 uses seven doctoral students familiar with personality/SJT development and 70 items: 21 translated human items plus seven LLM conditions of seven items. Study 2 uses another seven doctoral students from Beijing Normal University and 200 ChatGPT-5 items, 5 facets x 5 rounds x 8 items. Study 3 recruits 468 people in two samples; after excluding 25, 443 valid remain, 264 women (59.6 %), age 18-58, mean 28.25, and 97.5 % with a university degree or higher. The criterion subsample is 130 students, 67 women, mean age 21.12; the retest includes 80, 51 women, mean 31.60. The data confirm 233 main, 130 criterion, and 80 retest within the 443 IDs.

Findings

  • Temperature 1.0 and 1.1 tie on CVI 0.714; 1.0 obtains better ranks on need/options, not a single maximum of CVI.
  • Prompt v2 reaches CVI 0.755, but v1 achieves better overall mean quality, 6.57 versus 6.00, with no significant v1 to v2 improvement.
  • The 200 ChatGPT-5 items have global CVI 0.707; by facet 0.707/0.657/0.843/0.600/0.729 for N/E/O/A/C.
  • Fifty of 200 items fail CVR 0.71 and 59 fall below 6/7 in quality; 32 have scoring rationality <3.5/4.
  • No round effect is detected, but the test does not establish equivalence or reproduction of concrete forms.
  • The ANOVA of Study 2 is reproduced; only scoring rationality has a nominal facet effect, F(4,175)=2.833, p=0.026.
  • Study 3 reproduces alpha 0.745/0.844/0.705/0.568/0.607 and retest 0.578/0.770/0.414/0.521/0.649.
  • LLM-SJT/NEO convergence is 0.675/0.665/0.478/0.357/0.440 for N/E/O/A/C.
  • The CFA reports CFI/TLI 0.95 and RMSEA 0.03, but SRMR 0.11.
  • Eight of 35 LLM correlations with criteria are p<0.05; four survive 0.05/35.
  • The five open datasets have no missing cells, the IDs are unique, and criterion/retest link exactly with pretest.
  • There are no exact normalized duplicates in stems or options within the 310 entries.
  • The stored calculations of CVR, means, sums, alphas, correlations, and ANOVA are reproduced from raw data.
  • Study 1 and 3 used ai.shanxl.com, not an auditable official OpenAI call; the exact backend cannot be verified.
  • OSF publishes SPSS viewers, not syntax; Mplus inputs, not outputs; generation logs, time, tokens, and cost are missing.
  • All generated items fix A/B=high trait and C/D=low trait, a positional key not evaluated.

Limitations

  • The study generates personality instruments for humans; it does not study the synthetic personality of the model.
  • Each optimization condition uses only seven items and the same seven judges who determine the selection.
  • There is no preregistration, power analysis, optimization holdout, or equivalence intervals.
  • Non-significance is interpreted as stability/reproducibility although the design is not a test of equivalence.
  • The GPT-4 versus ChatGPT-5 contrast is not matched on facet, number of items, date, interface, or backend.
  • The third-party GPT-4 gateway prevents confirming identity, system prompt, and effective parameters.
  • ChatGPT-5 is identified by interface/date, not by an immutable model ID or conversation export.
  • The experts are seven doctoral students; the ICC of several indicators is only moderate.
  • Fifty Study 2 items fail CVR and 59 fail the quality cutoff, so expert review is needed.
  • Compliance and self-discipline have alpha 0.57/0.61; openness retest 0.41; compliance convergence 0.36.
  • SRMR 0.11 contradicts a reading of uniformly satisfactory fit.
  • No loadings, residuals, or Mplus outputs are published to audit the full CFA.
  • The mean of interfactor correlations is not a sufficient test of discrimination.
  • The 35 criterion correlations per instrument are interpreted without a prespecified correction for multiplicity.
  • The criteria are all Likert self-reports and do not include behavior, performance, or external informants.
  • A/B always codes high and C/D low; there is no randomization, position control, coaching, or faking.
  • Only one facet per domain is measured, not the full Big Five.
  • The sample is Chinese, highly educated, and partially student-based; there is no invariance or cross-cultural replication.
  • The new SJT is not directly compared with an equivalent Chinese expert SJT in Study 3.
  • Times/costs are estimates not backed by logs of the real experimental flow.
  • The release does not contain SPSS syntax, Mplus outputs, or a complete executable pipeline.
  • No seeds, system prompts, request IDs, failures, retries, or generation history are published.
  • Table S8 repeats Self-discipline where the first row corresponds to self-consciousness, a minor labeling error.

What the study does not establish

  • It does not demonstrate psychological personality, identity, consciousness, or stable traits of the LLM.
  • It does not confirm that each response actually came from gpt-4-1106-preview.
  • It does not demonstrate reproducibility across future ChatGPT versions, providers, or defaults.
  • It does not demonstrate temporal equivalence from p>0.05.
  • It does not demonstrate causal generalization across models with the available descriptive contrast.
  • It does not demonstrate satisfactory validity and reliability across the five facets.
  • It does not demonstrate discriminant validity just because the mean correlations are lower.
  • It does not demonstrate superiority or equivalence against a directly comparable expert SJT.
  • It does not demonstrate full Big Five coverage.
  • It does not demonstrate resistance to faking, coaching, social desirability, or position bias.
  • It does not demonstrate cultural adequacy outside of China or invariance across groups.
  • It does not demonstrate the cost or time declared in the actually used flow.
  • It does not allow end-to-end reproduction of SPSS/Mplus analysis and generation from OSF.

Traceability

Scope: Full text

Version: arXiv:2412.12144v4, revised 8 February 2026, 67 pages; Computers in Human Behavior Reports 21 (March 2026), DOI 10.1016/j.chbr.2026.100964; official OSF project jbvq7

Consulted source: https://arxiv.org/abs/2412.12144v4

Review: Codex complete bilingual fidelity pass using all 67 pages of arXiv v4, all-page visual inspection, publisher/DOI verification, all 16 OSF supplement pages, complete OSF item-bank and file-tree audit, raw .sav linkage and integrity checks, independent CVR/alpha/item-total/correlation/criterion/ANOVA recomputation, CFA-input inspection, model-provenance and reproducibility analysis; 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

  • Third-party gateway labelled gpt-4-1106-preview
  • Consumer ChatGPT instant mode labelled ChatGPT-5, 6 November 2025
  • DeepL for German-Chinese translation and back-translation

Instruments and metrics

  • Personality situational judgment tests
  • Content Validity Ratio and Content Validity Index
  • Option rationality, scoring rationality and overall item quality ratings
  • NEO-PI-R selected facets
  • Cronbach alpha and corrected item-total correlation
  • Two-week test-retest Pearson correlation
  • Five-factor WLSMV confirmatory factor analysis
  • Subjective well-being, depression, game addiction, aggression and Dark Triad criteria

Data used

  • OSF Study 1 expert-rating data, 70 items
  • OSF Study 2 expert-rating data, 200 items
  • OSF Study 3 pretest data, 443 participants
  • OSF Study 3 criterion data, 130 participants
  • OSF Study 3 retest data, 80 participants
  • OSF complete 310-item bank and CFA data

Evidence and location

  • Objectives, three-study design, and scope: arXiv v4 pages 1-14, abstract, introduction and Figure 1
  • Study 1 prompts, temperatures, judges, CVR/CVI and results: arXiv v4 pages 14-31 and Supplement pages 2-14
  • Study 2 ChatGPT-5 rounds, facets, ratings and factorial analyses: arXiv v4 pages 31-39 and Supplement pages 15-16
  • Study 3 sample, psychometrics, CFA and criterion validity: arXiv v4 pages 39-50, Tables 3-6
  • Interpretation, efficiency, cultural claims and limitations: arXiv v4 pages 50-57, Sections 7-8
  • Third-party GPT-4 access and effective parameter disclosure: Official OSF Supplementary Materials pages 12-13, Section 2 and Table S2
  • Item counts, duplicates and fixed scoring positions: Official OSF Item Bank.xlsx; independent audit of all Study1/Study2/Study3 sheets on 16 July 2026
  • Raw-data integrity and independent statistical recomputation: Official OSF Study1/Study2/Study3 .sav files; independent CVR, alpha, item-total, Pearson, criterion and balanced-ANOVA calculations
  • CFA and missing executable analysis artifacts: Official OSF CFA_data_443.dat, GPSJT.inp, NEO-PI-R.inp and complete OSF file-tree audit
  • Publication and version history: Computers in Human Behavior Reports DOI 10.1016/j.chbr.2026.100964 and arXiv:2412.12144v4 submission history
  • Integral psychometric and artifact audit: reports/verification/article-204-sjt-generation-psychometrics-and-artifact-audit.json
  • Complete visual inspection: All 67 manuscript pages and all 16 OSF supplement pages rendered and visually inspected on 16 July 2026