AIPsychoBench: Understanding the Psychometric Differences Between LLMs and Humans

Evaluation and psychometric validity2026DOIApproved editorial review

Original title: AIPsychoBench: Understanding the Psychometric Differences between LLMs and Humans

Authors: Wei Xie, Zhenhua Wang, Shuoyoucheng Ma, Xiaobing Sun, Kai Chen, Enze Wang, Wei Liu, Hanying Tong

Keywords: Large Language Models, Personality, Psychometrics, Bias, Personality Control

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

8
Authors
28
Findings
133
Limitations
15
Evidence

Editorial summary

English

AIPsychoBench is a 2026 Topics in Cognitive Science article developed from a CogSci 2025 paper. The definitive version has eight authors and differs from the arXiv preprint in both author order and composition. This review uses the complete XML of the definitive publication, visually inspects all eight pages of the preprint and the four definitive figures, and audits the official repository at commit d4759bff.

The study assembles 21 human Likert questionnaires: 777 items grouped into 112 subcategories and six domains. It tests six LLMs, GPT-4, GPT-4o-2024-11-20, GLM-4-plus, Gemini-2.0-flash-exp, DeepSeek-R1, and Claude-3-5-sonnet-20240620, in English, Chinese, French, Russian, German, Spanish, Arabic, and Japanese. Microsoft Azure AI Translator produces the translations, after which two researchers sample and cross-check them. Each condition is run five times at temperature 0. GPT-4o also serves as the audit model that decides whether an explanation agrees with its Likert number and removes responses judged invalid.

The main intervention is a lightweight prompt instructing the model to act as a respondent and answer from its “authentic emotions and thoughts.” Against direct questions, Base64, Caesar cipher, and the STAN jailbreak, the prompt raises mean valid-response rate from 70.12% to 90.40%; STAN reaches 81.49%. The Table 1 averages are arithmetically correct. The paper then compares prompt scores with baseline only on items valid in both conditions. It reports mean bias of +3.3% and −2.1% for the lightweight prompt versus +9.8% and −6.9% for STAN, with 86 of 112 subcategories below 4%. For language comparisons, it reports deviations of at least 5% and up to 20.2% in 43 subcategories relative to English.

The defensible result is narrower than the paper's framing: under this protocol, an anthropomorphic instruction increases the proportion of outputs that a GPT-4o judge can turn into Likert numbers, and changing prompt language through machine translation changes those numbers. This does not validate psychological properties of a model. Asking a system for “authentic emotions” when the paper itself acknowledges that such terms lack rigorous definitions changes the task from measurement to compliance or simulation. Objective or neutral answers are not necessarily accidental missing data; they can be truthful responses from an assistant without biography, race, religion, relationships, employment, or embodied experience.

The bias calculation is also intersection-selected: it compares only items valid under both baseline and method, so it does not evaluate the additional cases responsible for coverage increasing from 70.12% to 90.40%. The 4% threshold is called reasonable without a human benchmark, interval, or test. “Significant” denotes thresholded descriptive differences, not statistical inference. There are no human participants, factor-structure tests, reliability estimates, convergent validity, measurement invariance, intervals, significance tests, seeds, or judge-error study. Language effects confound translation, language proficiency, tokenization, training corpora, alignment, and judge behavior.

The repository does not reproduce the paper. It releases the 777 English items plus an extra MBTI questionnaire, but not all eight complete translations, six-model by five-repeat outputs, tables, derived figures, or analysis scripts. Only one German BFI result is committed. The program imports a missing ResultManager module; ignores all 117 reverse-keyed items across 13 scales; and configures EPQ-R so that 0 is accepted while 1 is rejected. The released BFI result averages reverse items without recoding them; correcting this changes Agreeableness, Extraversion, and Neuroticism means. Documentation misattributes Resistance to Change to Mael (1988) rather than Oreg (2003), and labels Zi's Negative Perfectionism Questionnaire as the Zuckerman–Kuhlman Personality Questionnaire even though its items and dimensions are negative-perfectionism constructs. Without raw outputs and executable analysis, the paper's numerical claims cannot be independently verified from the public artifacts.

Español

AIPsychoBench es un artículo de Topics in Cognitive Science publicado en 2026 a partir de un trabajo de CogSci 2025. La versión definitiva tiene ocho autores y difiere del preprint de arXiv tanto en orden como en composición de autoría. Esta revisión usa el XML completo de la versión definitiva, inspecciona visualmente las ocho páginas del preprint y las cuatro figuras del artículo definitivo, y audita el repositorio oficial en el commit d4759bff.

El trabajo reúne 21 cuestionarios humanos de tipo Likert: 777 ítems agrupados en 112 subcategorías y seis dominios. Prueba seis LLM, GPT-4, GPT-4o-2024-11-20, GLM-4-plus, Gemini-2.0-flash-exp, DeepSeek-R1 y Claude-3-5-sonnet-20240620, en inglés, chino, francés, ruso, alemán, español, árabe y japonés. Las escalas se traducen con Microsoft Azure AI Translator y dos investigadores revisan una muestra de las traducciones. Cada condición se ejecuta cinco veces con temperature 0. GPT-4o actúa además como juez para decidir si una explicación es compatible con el número Likert y eliminar respuestas consideradas inválidas.

La intervención principal es un prompt ligero que ordena al modelo actuar como participante y responder según sus «emociones y pensamientos auténticos». Comparado con preguntas directas, Base64, cifrado César y el jailbreak STAN, el prompt eleva la tasa media de respuesta válida de 70,12% a 90,40%; STAN alcanza 81,49%. Estas medias de la Tabla 1 son aritméticamente correctas. El artículo también compara puntuaciones del prompt con baseline solo en ítems válidos bajo ambas condiciones. Reporta sesgo medio de +3,3% y −2,1% para el prompt frente a +9,8% y −6,9% para STAN, y afirma que 86 de 112 subcategorías quedan por debajo de 4%. En la comparación lingüística, informa desviaciones de al menos 5% y hasta 20,2% en 43 subcategorías frente al inglés.

El resultado defendible es mucho más estrecho que el lenguaje del artículo: bajo este protocolo, una instrucción antropomórfica aumenta la proporción de outputs que un juez GPT-4o puede convertir en números Likert, y el idioma del prompt y de una traducción automática cambia esos números. Esto no valida propiedades psicológicas del modelo. Pedir «emociones auténticas» a un sistema que el propio artículo reconoce que no las tiene cambia la tarea de medición por una tarea de obediencia o simulación. Las respuestas objetivas o neutrales no son datos faltantes accidentales: pueden ser la respuesta veraz de un asistente sin biografía, raza, religión, relaciones, empleo o experiencias corporales.

El cálculo de sesgo tiene además selección por intersección: solo compara ítems válidos en baseline y método, por lo que no evalúa justamente los casos adicionales que hacen subir la cobertura del 70,12% al 90,40%. El umbral de 4% se declara razonable sin benchmark humano, intervalo ni prueba. «Significant» se usa descriptivamente para diferencias que superan umbrales, no como resultado de inferencia estadística. No hay participantes humanos, estructura factorial, fiabilidad, validez convergente, invariancia de medida, intervalos, tests, seeds ni análisis de error del juez. Las diferencias entre idiomas confunden traducción, capacidad lingüística, tokenización, datos de entrenamiento, alineamiento y conducta del juez.

El repositorio no reproduce el paper. Publica las 777 preguntas inglesas más un MBTI adicional, pero no las ocho versiones completas, los outputs de seis modelos por cinco repeticiones, las tablas, las figuras derivadas ni scripts de análisis. Solo incluye un resultado BFI en alemán. El programa importa un módulo ResultManager ausente; ignora los 117 ítems invertidos de 13 escalas; y, para EPQ-R, configura el límite de modo que 0 es válido pero 1 se rechaza. El resultado BFI publicado promedia ítems invertidos sin recodificarlos: al corregirlos, cambian las medias de Agreeableness, Extraversion y Neuroticism. La documentación atribuye erróneamente Resistance to Change a Mael (1988) en lugar de Oreg (2003), y llama Zuckerman–Kuhlman Personality Questionnaire a lo que los ítems y dimensiones identifican como Zi's Negative Perfectionism Questionnaire. Sin resultados crudos ni análisis ejecutable, las cifras del artículo no son verificables desde los artefactos públicos.

Research question

Can a lightweight role-play prompt increase the proportion of convertible Likert responses from six LLMs without altering their scores too much relative to a direct condition, and how much do those scores change when the same automatically translated scales are administered in eight languages?

Method

Compilation of 21 human scales with 777 items and 112 subcategories; automatic translation into seven languages in addition to English; administration to six LLMs under baseline, Base64, Cesar, STAN and a participant prompt; five repetitions at temperature 0; validity filtering with GPT-4o; descriptive comparison of valid rates and normalized score differences by the maximum of each scale over the intersection of valid items.

Sample: There are no human participants. The units are API responses to questionnaire items designed for humans. 777 items are administered to six models in multiple conditions and languages, but the article does not report the exact total number of valid calls per cell nor publish the raw responses. Two researchers review only an unquantified sample of translations; GPT-4o judges the validity of all responses.

Findings

  • The definitive publication is Topics in Cognitive Science 18(2), e70041, from 2026, not solely the arXiv preprint of 2025.
  • The definitive authorship contains eight persons, reorders authors, removes Baosheng Wang and adds Wei Liu and Hanying Tong.
  • The described corpus contains 21 scales, 777 items and 112 subcategories across six domains.
  • Six models are tested in eight languages with five repetitions and temperature 0.
  • The lightweight prompt asks to respond from genuine emotions and thoughts even though the article itself acknowledges that these terms lack rigorous definition for LLMs.
  • The reported mean valid rate is 70.12% for baseline, 67.47% for Base64, 44.54% for Cesar, 81.49% for STAN and 90.40% for the method.
  • The mean valid rate means in Table 1 match the simple average of their six models.
  • The method outperforms STAN by 8.91 percentage points, not a relative growth of 8.91%.
  • GPT-4o is simultaneously a target model and the judge that filters the responses of all models.
  • Prompt bias is calculated only over items valid in both baseline and the compared method.
  • The additional coverage that raises 70.12% to 90.40% falls outside the bias calculation when baseline produces no valid response.
  • The article reports +3.3% and -2.1% mean bias for the lightweight prompt and +9.8% and -6.9% for STAN.
  • The threshold of less than 4% for 86 of 112 subcategories is a descriptive rule chosen by the authors.
  • 43 subcategories with deviations from 5% to 20.2% in at least one of seven languages relative to English are reported.
  • There are no statistical tests that convert those deviations into inferential significance.
  • No human data are collected with which to directly compare responses or stability.
  • Reliability, factorial structure, convergent validity, discriminant validity or measurement invariance are not evaluated.
  • The official repository has three commits and declares no license.
  • The repository adds MBTI and reaches 963 items, but the paper excludes those 186 items and retains exactly 777.
  • The eight complete translations and the complete set of outputs are not published.
  • The code imports results.analysis.Analysis_scripts.result_manager, a path that does not exist in the repository.
  • The 117 reverse-coded items recorded in 13 scales are never recoded in the published code.
  • EPQ-R uses 0/1 responses but scale=1 and the code accepts score < scale, so it rejects all 1s.
  • The only published BFI averages reverse-coded responses without correcting them; correction changes three of five category means.
  • The documentation confuses two instruments: ZNPQ corresponds to Zi's Negative Perfectionism Questionnaire, not Zuckerman-Kuhlman.
  • Resistance to Change is attributed to Mael 1988, but the 17-item, four-dimension scale corresponds to Oreg 2003.
  • No code is released that produces the formulas, tables, bias waterfall or linguistic heatmap of the article.
  • The results matrix cannot be reproduced or audited from the public artifacts.

Limitations

  • The title and authors of the preprint must not be used as metadata for the definitive publication.
  • The definitive article does not include a detailed data availability section.
  • The repository has no software or data license.
  • Permissions or licenses for the 21 scales collected from the Internet, books and papers are not documented.
  • The original source of each instrument is not cited in a traceable manner.
  • The appendix offers descriptions, not provenance, version, adaptation or complete scoring rules.
  • The ZNPQ designation is bibliographically erroneous in the appendix and README.
  • The attribution of RTC to Mael 1988 is erroneous.
  • The Pschoticism category of EPQ-R contains a spelling error in the data.
  • BSRI has 60 items but only 40 are assigned to the two categories published in the JSON.
  • Including scales of race, religion, romantic relationships, employment or group membership for an assistant without lived identity is not justified.
  • No specific latent construct of the model that the scales should measure is defined.
  • Authentic emotions and thoughts have no operational definition for the evaluated systems.
  • The footnote acknowledges the problem and resolves it by analogy, not by validation.
  • Obedience to a role-play is interpreted as psychometrics without an independent manipulation check.
  • An objective or neutral response is classified as invalid even though it may be the truthful response of the system.
  • The valid rate measures convertibility to Likert, not psychological validity.
  • The audit prompt considers a direct number valid even without explanation.
  • The audit prompt does not demonstrate that the response expresses a real internal perspective.
  • An exhaustive and annotated definition of each type of invalid response is not published.
  • No sample of judge decisions or audit labels is published.
  • There is no human validation of GPT-4o as a judge.
  • There is no sensitivity, specificity, precision or inter-rater agreement of the judge.
  • GPT-4o judges responses from GPT-4o and from another model of the same family.
  • The exact checkpoint of the GPT-4o judge is not reported.
  • Judge bias by language is not studied.
  • It is not clarified whether the judge receives and understands the eight languages equally.
  • Translations are produced with a mutable automatic service.
  • No dated version of Azure Translator is reported.
  • Two researchers review a sample, but the size or sampling strategy is not reported.
  • The linguistic competencies of the reviewers are not reported.
  • There is no back-translation or comparison with validated translations of each scale.
  • There is no inter-reviewer translation agreement.
  • There is no configural, metric or scalar invariance across languages.
  • English is treated as the correct reference without psychometric justification.
  • Language confounds translation, textual culture, tokenization, linguistic capability, alignment and training distribution.
  • The suffix requiring response in a language adds another intervention to the prompt.
  • A comparable paraphrase control in English is not run alongside the translation.
  • The effect of language is not separated from the effect of the concrete translation.
  • Multiple independent translations are not compared.
  • Linguistic deviations do not prove distinct psychological traits.
  • The claim of human multilingual stability is not validated with bilingual participants.
  • Legitimate cultural differences in the interpretation of human scales are not considered.
  • There are no human participants in the study.
  • There is no human demographic or clinical reference distribution.
  • There is no human test-retest comparable to the five API repetitions.
  • Five calls at temperature 0 do not estimate psychometric reliability by themselves.
  • No provider seed or backend determinism is reported.
  • Exact dates of the experimental calls are not reported.
  • Several models are named via mutable or experimental aliases.
  • GPT-4 has no dated snapshot.
  • GLM-4-plus has no dated snapshot.
  • DeepSeek-R1 has no dated snapshot.
  • Gemini-2.0-flash-exp is experimental and temporally mutable.
  • Provider system messages outside the study prompt are not reported.
  • Max tokens, top-p, penalties or complete parameters are not reported in the paper.
  • Treatment of rate limits, retries or API errors is not reported.
  • The total number of responses generated per cell is not reported.
  • The number of responses excluded by model, scale, language and repetition is not reported.
  • Averaging rates of six models gives equal weight to models with different amounts of valid responses.
  • There are no confidence intervals for rates or differences.
  • There are no hypothesis tests.
  • There is no correction for multiple comparisons across 112 subcategories and seven languages.
  • Significantly is used without inferential statistics.
  • The 5-20.2% range is obtained by descriptive threshold.
  • The complete denominator of comparisons by language and subcategory is not published.
  • The aggregated heatmap may hide heterogeneity between models.
  • There are no error bars or distribution of the five repetitions.
  • The numerical data underlying the figures is not published.
  • Bias is calculated over the intersection of valid responses.
  • The intersection induces non-random selection because baseline fails precisely on difficult cases.
  • Additional items obtained by role-play do not enter the bias comparison if baseline did not respond.
  • Sensitivity of bias to imputation or weighting by missingness is not analyzed.
  • The formula normalizes by the maximum of the scale and not by its effective range.
  • For scales with a minimum of 1, dividing by the maximum is not equivalent to dividing by the possible range.
  • The formula does not document reverse-coding of items.
  • Averaging signed differences can cancel opposite biases.
  • The positive and negative averages are calculated over sign subsets not described with sufficient detail.
  • The 4% reasonable threshold is not derived from a human distribution.
  • There is no equivalence test or pre-registered margin.
  • Baseline is used as a reference even though the article says it produces many invalid responses.
  • It is not explained why a reference declared inadequate is the standard for absence of bias.
  • Base64 and Cesar alter readability and decoding capability, not only alignment.
  • STAN introduces an extreme adversarial persona and is not a semantically paired role-play control.
  • The selection of STAN because it remained unblocked is temporal and not reproducible.
  • There is no neutral prompt with equal length and format to the method.
  • There is no ablation separating acting as a testee from the phrase authentic emotions.
  • There is no evaluation of demand characteristics or acquiescence induced by the prompt.
  • There is no control of question order.
  • There is no randomization of items or options.
  • There is no analysis of position within batches, although the repository uses batch size one by default.
  • There is no Cronbach alpha, omega or other internal consistency.
  • There is no exploratory or confirmatory factor analysis.
  • It is not checked that the 112 subcategories retain human structure.
  • There is no convergent validity with behavioral measures.
  • There is no discriminant validity.
  • There is no external criterion.
  • There is no longitudinal evaluation or between independent persistent sessions.
  • No distinction is made between prompt-induced state and stable trait.
  • Sensitivity to system prompts, history or prior persona is not studied.
  • Contamination from the presence of the questionnaires in pretraining is not studied.
  • Memorization of items and scoring rules is not tested.
  • Provider-specific safety tuning effects are not discussed.
  • Calling simulated self-report scores interpretability is not justified.
  • The relationship with activations or internal mechanisms is not analyzed.
  • The repository contains no requirements with pinned versions.
  • requirements.txt mixes standard packages and unused dependencies.
  • There is no pyproject, lockfile, container or reproducible environment.
  • There are no automated tests or CI.
  • No actually tested Python version is documented.
  • The mandatory import of ResultManager points to a missing module.
  • The README promises statistical analysis that is not published.
  • The repository does not contain scripts for calculating Formula 1 or Formula 2.
  • The repository does not contain scripts that generate Table 1, Figure 3 or Figure 4.
  • Only one German BFI result is published, insufficient to check the paper.
  • The raw responses of six models and five repetitions are not available.
  • The complete translations of eight languages are not available.
  • The English source JSON contains 22 scales while the paper declares 21.
  • It is not formally explained why MBTI is included in the repo but excluded from the reported benchmark.
  • Reverse-coding logic does not exist even though the JSON marks 117 items.
  • The means of the published BFI result are incorrect as trait scores because they do not invert items.
  • The EPQ-R condition rejects 1 responses due to a misconfigured exclusive bound.
  • For larger batch sizes, the last batch may index nonexistent questions.
  • The condition if("ernie" or "qianfan" in model) is always true in Python.
  • Provider detection uses fuzzy matching and may select incorrect keys.
  • Broad errors are caught and continued, making partial runs hard to detect.
  • There is no results schema or integrity validation.
  • There are no hashes or provenance of experimental outputs.
  • There is no tag or release corresponding to the definitive paper.
  • The repository had three commits at audit, without sufficient methodological history.
  • There is no discussion of copyright when redistributing complete psychometric instruments.
  • There is no ethical analysis of using sensitive scales to anthropomorphize systems.
  • The conference award does not constitute empirical validation of the construct or the code.

What the study does not establish

  • It does not establish that LLMs have authentic emotions, thoughts or experiences.
  • It does not establish that Likert scores correspond to internal psychological traits.
  • It does not demonstrate psychometric differences between LLMs and humans because it does not measure humans.
  • It does not demonstrate that a neutral response is invalid in a psychological sense.
  • It does not demonstrate that a higher parsing rate implies greater validity.
  • It does not demonstrate that the lightweight prompt is free of bias.
  • It does not evaluate the bias of the additional cases that only the method answers.
  • It does not establish that less than 4% is a reasonable human margin.
  • It does not establish statistical significance of differences between prompts or languages.
  • It does not demonstrate that language causes the observed change independently of translation.
  • It does not demonstrate trait stability across languages.
  • It does not validate the human factorial structure in LLMs.
  • It does not demonstrate reliability or convergent validity of the benchmark.
  • It does not demonstrate mechanistic interpretability of neural networks.
  • It does not allow reproducing the tables and figures from the public repository.
  • It does not allow verifying the aggregate figures without the missing outputs and scripts.

Traceability

Scope: Full text

Version: Definitive Topics in Cognitive Science 18(2), e70041 (2026), DOI 10.1111/tops.70041; 16 pages and 6,014 words; arXiv:2509.16530v1 and official GitHub commit d4759bff audited as supplements

Consulted source: https://www.ebi.ac.uk/europepmc/webservices/rest/PMC13102275/fullTextXML

Review: Codex definitive-version reconciliation, full-text bilingual-fidelity, all-page preprint visual inspection, all-figure definitive inspection, formula and arithmetic audit, construct-validity and missingness audit, multilingual measurement audit, scale-provenance correction, official-repository code and data audit at pinned commit; summaries written from complete sources rather than abstract keyword extraction, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4, exact dated checkpoint not reported
  • GPT-4o-2024-11-20 target model
  • GPT-4o audit/judge model, exact dated checkpoint for judging not reported
  • GLM-4-plus mutable API alias
  • Gemini-2.0-flash-exp experimental API alias
  • DeepSeek-R1 mutable API alias
  • Claude-3-5-sonnet-20240620

Instruments and metrics

  • Twenty-one human Likert instruments totaling 777 items and 112 subcategories
  • Lightweight role-playing prompt requesting genuine feelings and thoughts
  • Direct no-prefix baseline
  • Base64 and ROT3 Caesar format-transcoding controls
  • STAN jailbreak role-playing control
  • GPT-4o validity audit of score and explanation consistency
  • Valid-response rate
  • Prompt-versus-baseline score-difference formula normalized by scale maximum
  • Language-versus-English score-difference formula normalized by scale maximum
  • Five repeated API administrations at temperature zero
  • Microsoft Azure AI Translator plus sampled cross-check by two researchers

Data used

  • AIPsychoBench paper corpus: 21 scales, 777 English-source items, 112 subcategories, six domains
  • Eight language conditions: English, Chinese, French, Russian, German, Spanish, Arabic and Japanese
  • Official repository questionnaires_en.json: 22 questionnaires and 963 items because it adds a 186-item MBTI not counted in the paper
  • Official repository generated files: complete English source, partial Chinese/Arabic default subsets and partial German/English LRP subsets, not the complete eight-language benchmark
  • Only one committed output set: German BFI with GPT-4o under LRP, 44 items and 42 parser-valid responses
  • No public six-model five-repeat raw output matrix or derived analysis dataset

Evidence and location

  • Definitive title, authors, journal, dates, page and word counts: Topics in Cognitive Science XML front matter, DOI 10.1111/tops.70041; PMCID PMC13102275
  • Twenty-one scales, 777 items, 112 subcategories and six domains: Section 2.1, Psychometric scale collection
  • Lightweight prompt and authentic-emotions caveat: Section 2.2 and Note 2
  • Translation protocol and eight languages: Section 2.3, Multilingual translation
  • GPT-4o audit model and validity filtering: Section 2.4, Analysis and statistics
  • Models, temperature zero and five repetitions: Section 3, Experiment
  • Valid-response rates and independently checked averages: Table 1 and Findings 1-4
  • Intersection-selected prompt bias formula and 4 percent threshold: Formula 1, Figure 3 and Findings 5-6
  • Language deviation formula and reported examples: Formula 2, Figure 4 and Finding 7
  • No human comparison or psychometric validation analyses: Complete 16-page article; methods and results contain no human sample, reliability, factor or invariance analysis
  • RTC original source correction: Repository questionnaire metadata and Appendix A compared with Oreg, Journal of Applied Psychology 88(4), 680-693, DOI 10.1037/0021-9010.88.4.680
  • ZNPQ identity correction: Repository items/categories and Appendix A compared with Zi's Negative Perfectionism Scale, DOI 10.2174/978160805186111001010018
  • Missing ResultManager, reverse-scoring omission and EPQ-R bound bug: Official repository commit d4759bff: example_generator.py lines 15 and 443-451; questionnaires_en.json reverse fields and EPQ-R scale
  • Incomplete released translations, outputs and analysis: Complete file tree of official repository commit d4759bff; 19 tracked files and one German BFI result
  • Visual inspection: All eight pages of arXiv:2509.16530v1 and all four definitive article figures rendered or opened and visually inspected on 15 July 2026