Is Machine Psychology here? On Requirements for Using Human Psychological Tests on Large Language Models

Reviews, theory, and governance2024ACL AnthologyApproved editorial review

Original title: Is Machine Psychology Here? On the Requirements for Using Human Psychological Tests on LLMs

Authors: Lea Löhn, Niklas Kiehne, Alexander Ljapunov, Wolf-Tilo Balke

Keywords: Large Language Models, Psychological Assessment, Machine Psychology, Test Reliability, Construct Validity

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

4
Authors
13
Findings
19
Limitations
15
Evidence

Editorial summary

English

Löhn, Kiehne, Ljapunov, and Balke do not run a new personality test on models. They construct a normative framework for deciding whether human psychological tests can legitimately be transferred to LLMs. Drawing on the Standards for Educational and Psychological Testing and the International Guidelines for Test Use, they formulate seven requirements: reliability for the intended use (R1), validity for that use (R2), suitability for the test taker (R3), non-disclosure or contamination control for test materials (R4), validity for every model being compared (R5a), validity of translations (R5b), and transparent, reproducible test use (R5c). The central contribution is to separate repeatable responding from justified score interpretation: a system may always return the same answer while still failing to measure the human construct attributed to it.

To illustrate the framework, the authors review 25 studies found through keyword searches in Google Scholar, Scopus, and DBLP up to October 2023, followed by citation-network tracing. The sample covers 12 constructs and 34 tests or assessments. Twenty studies include GPT-3 or a later version, and 17 of those 20 examine no other model family. All four authors jointly rate each paper on an ordinal scale: not applicable, not addressed, discussed, an appropriate effort or study without supporting evidence, and any evidence of fulfillment. The rule is intentionally lenient: the top category recognizes some evidence for one form of reliability or validity rather than complete validation, and evidence sufficiency is not judged.

The table shows a substantial evidence gap. For R1, only 6 of 25 studies provide any reliability evidence; for R2, only 3 provide any validity evidence. R3 appears better covered, but its 13 top-rated cases benefit from a weak operational rule that treats a compatible text format as suitable. Only one study reaches the top category for R4. Among the ten comparative studies to which R5a applies, only Serapio-García et al. provides evidence. R5b applies to two studies, and only Pellert et al. uses already validated translations. Just 2 of 25 fully meet R5c. Nine studies modify or rephrase tests to reduce contamination, but only Coda-Forno et al. supplies evidence that the modification preserves results. No study provides evidence for every applicable requirement.

The diagnosis is useful as a checklist and as a warning against anthropomorphizing scores without a validation chain. It is not, however, a reproducible systematic review or a validated standard. The paper releases no search strings, selection flow, exclusion criteria, coding sheet, evidence excerpts, coder agreement, or artifact; ratings are decided in joint meetings. It also supplies no operational threshold for how much evidence is enough. Some requirements are overly absolute or narrow: proving absence of contamination is often impossible for proprietary models, suitability is not equivalent to accepting text, and the fairness section omits subgroup bias, accessibility, and harm. The paper establishes that the reviewed early literature incompletely documented its inferences under this framework. It does not establish that every substantive conclusion was false or that human constructs can never be adapted through model-specific validation.

Español

Löhn, Kiehne, Ljapunov y Balke no ejecutan un nuevo test de personalidad sobre modelos: construyen un marco normativo para juzgar si es legítimo trasladar tests psicológicos humanos a LLM. A partir de los Standards for Educational and Psychological Testing y las International Guidelines for Test Use formulan siete requisitos: fiabilidad para el uso previsto (R1), validez para ese uso (R2), adecuación al tipo de examinado (R3), no divulgación o contaminación del material (R4), validez en cada modelo comparado (R5a), validez de las traducciones (R5b) y uso transparente y reproducible del test (R5c). El mérito central es separar obtener una respuesta repetible de justificar la interpretación del score: un sistema puede contestar siempre igual y aun así no medir el constructo humano que se le atribuye.

Para mostrar la utilidad del marco, revisan 25 trabajos localizados mediante búsquedas por palabras clave en Google Scholar, Scopus y DBLP hasta octubre de 2023, ampliadas siguiendo redes de citas. El conjunto cubre 12 constructos y 34 tests o evaluaciones; 20 estudios incluyen GPT-3 o una versión posterior y 17 de esos 20 no estudian otra familia. Los cuatro autores califican conjuntamente cada paper con una escala ordinal: no aplicable, no abordado, discutido, esfuerzo o estudio sin evidencia de cumplimiento, y cualquier evidencia de cumplimiento. La propia regla es deliberadamente benévola: la categoría superior reconoce alguna evidencia de una forma de fiabilidad o validez, no una validación completa, y no valora si esa evidencia es suficiente.

La tabla revela una brecha amplia. Para R1, solo 6 de 25 estudios aportan alguna evidencia de fiabilidad; para R2, 3 aportan alguna evidencia de validez. R3 parece mejor atendido, pero los 13 casos superiores se benefician de una definición débil que considera adecuado un formato textual compatible. Solo 1 estudio alcanza la categoría superior en R4. De los 10 trabajos comparativos a los que se aplica R5a, solo Serapio-García et al. presenta evidencia; R5b se aplica a dos trabajos y solo Pellert et al. utiliza traducciones ya validadas. Únicamente 2 de 25 cumplen plenamente el criterio de transparencia R5c. Nueve estudios modifican o reformulan tests para reducir contaminación, pero solo Coda-Forno et al. aporta evidencia de que la modificación conserva resultados. Ningún trabajo aporta evidencia para todos sus requisitos aplicables.

El diagnóstico es valioso como lista de control y como advertencia contra antropomorfizar scores sin una cadena de validación. No es, sin embargo, una revisión sistemática reproducible ni un estándar validado. No se publican cadenas de búsqueda, flujo de selección, criterios de exclusión, hoja de codificación, extractos de evidencia, acuerdo entre evaluadores o artefactos; la clasificación se decide en reuniones conjuntas. Tampoco hay umbrales operativos para decidir cuánta evidencia basta. Algunas exigencias son demasiado absolutas o estrechas: demostrar ausencia de contaminación suele ser imposible en modelos propietarios, la «adecuación» no se reduce a aceptar texto y la sección de fairness omite sesgo entre grupos, accesibilidad y daño. El artículo establece que la literatura temprana revisada documentaba insuficientemente sus inferencias bajo este marco; no prueba que todas sus conclusiones fueran falsas ni que los constructos humanos no puedan adaptarse mediante validación específica.

Research question

What requirements derived from psychological assessment theory and standards should be met when applying human tests to LLMs, and to what extent does a sample of 25 early machine psychology studies satisfy them?

Method

Conceptual work with a demonstration narrative review. The authors derive R1–R5c from two normative bodies of assessment and test use, search literature via keywords in Google Scholar, Scopus and DBLP until October 2023, expand the set following citations and retain 25 papers. The four authors jointly assess each applicable requirement in meetings and publish an ordinal matrix per study. The editorial audit read and rendered the 13 pages, reconstructed the counts of Table 1, verified Tables 2–3 and searched without finding a coding sheet, data or public code associated.

Sample: The unit of analysis is the paper, not a person or a new model output. The sample contains 25 publications from June 2022 to September 2023 according to the text, although the final table incorporates Pellert et al. (2024). It covers cognition, personality, values, morality, decision biases, well-being and anxiety. Twenty works include GPT-3 or later and 17 of them do not compare another family. Requirements R5a and R5b are only scored when there are several models or languages, respectively.

Findings

  • The seven proposed requirements are R1 reliability, R2 validity, R3 adequacy, R4 disclosure/contamination control, R5a validity by model, R5b translation validity and R5c transparency of test use.
  • The editorially reconstructed matrix gives for R1: 6 with evidence, 6 with effort without evidence, 7 discussed and 6 not addressed.
  • For R2: 3 with evidence, 4 with study without sufficient evidence, 13 discussed and 5 not addressed.
  • For R3: 13 with evidence and 12 with effort without evidence; the result depends on considering a textual input/output format compatible.
  • For R4: 1 with evidence, 8 with effort without evidence, 7 discussed and 9 not addressed.
  • R5a does not apply to 15 studies; among the 10 remaining there is 1 with evidence, 1 with study without evidence, 1 discussed and 7 not addressed.
  • R5b does not apply to 23 studies; Pellert et al. receives evidence for using validated translations and Stevenson et al. remains as discussed for not sufficiently documenting the translation.
  • For R5c: 2 with evidence, 19 with effort without full compliance, 2 discussed and 2 not addressed.
  • None of the 25 studies obtains evidence in all the requirements that are applicable to it, even with the lenient scoring rule.
  • Nine of 25 modify or generate material to reduce contamination; only Coda-Forno et al. provides evidence of correspondence between original and reformulated items.
  • Of the 10 comparative studies to which R5a applies, only Serapio-García et al. offers evidence of validity by model and observes better results in larger and instruction-tuned models.
  • Only two works satisfy R5c because unfixed proprietary versions prevent reproducing a configuration even if parameters or code are published.
  • The paper concludes that there is no consensual methodology in the sample and that specific constructs and validation processes for LLMs are needed.

Limitations

  • It is a narrative and proof-of-concept review, not a systematic review: it does not publish exact strings, dates per database, protocol, complete inclusion/exclusion criteria, flow, duplicates or list of discards.
  • The search closes in October 2023 and covers a very early stage; it does not represent the state of the field in 2024 or in 2026.
  • The declared range June 2022–September 2023 does not fit literally with Pellert et al. (2024) in the final table, probably an updated version of a preprint, without explaining the bibliographic cutoff.
  • The assessments are decided collectively in meetings; there is no independent coding, blinding, inter-rater agreement, registry of disagreements or reproducible adjudication.
  • The extraction sheet, justification per cell, evidence citations, data, code or supplement is not published; only the visible matrix can be reconstructed.
  • The superior category means any evidence of a single form, not sufficient or complete validation; a reader may overinterpret the positive symbol.
  • The seven requirements are not validated via Delphi, external experts, user study or proof that they discriminate research with reliable conclusions.
  • There are no operational criteria for approval, evidence thresholds, sample sizes, order of priority or rules for transferring validity between versions of a model.
  • R3 is operationalized too weakly: that a test accepts text does not prove adequacy of content, response process, construct or interpretation for an LLM.
  • R4 formulates as a requirement that the training data does not contain the material, something normally not verifiable in proprietary models; it would be more operational to estimate risk, effect and mitigations of contamination.
  • The grouping under fairness is limited to validity between models, translation and transparency; it does not evaluate bias by groups, adverse impact, accessibility, cultural differences, DIF or harms.
  • R5c is mainly a transparency and reproducibility requirement, not a direct measure of fairness.
  • The claim that test–retest, alternate forms and internal consistency can be applied independently of the nature of the examinee simplifies assumptions about unit, population, stability and meaning of the items.
  • A deterministic generation only guarantees repeatability if the entire system remains fixed; hosted APIs, weights, kernels or infrastructure may change, and repeating an output does not equate to measuring well.
  • The ordinal matrix does not weight consequences, evidence quality, scope or number of instruments; all papers and requirements count equally.
  • The sample mixes very different constructs, paradigms and types of publication, so that the aggregate counts are descriptive and not a population estimate.
  • Possible conflicts, overlap among reviewed authors, citation bias or dependence introduced by citation-network tracing are not analyzed.
  • Requiring new validity for each model is prudent, but the work does not define when a family, update or configuration constitutes a new examinee or what evidence can be transferred.
  • The framework recognizes that it is not exhaustive and leaves unresolved constructs specific to LLMs, unknown response processes and the ambiguity between individual and population.

What the study does not establish

  • It does not demonstrate that the substantive conclusions of the 25 studies are false; it shows that their measurement support is incomplete according to this rubric.
  • It does not demonstrate that an LLM possesses or lacks personality, cognition, anxiety, values or theory of mind.
  • It does not prove that human psychological tests can never be adapted to models; it requires specific redefinition and validation.
  • It does not constitute an exhaustive review of all available literature or a current snapshot of the field.
  • It does not empirically validate that R1–R5c are necessary and sufficient for all machine psychology applications.
  • It does not offer a complete standard, certification, checklist with pass/fail or reproducible audit procedure.
  • It does not allow quantitatively comparing the global quality of the 25 papers by summing ordinal symbols.
  • It does not demonstrate absence or presence of contamination in any specific proprietary model.
  • It does not establish that a deterministic response is psychometrically reliable beyond its repeatability under a fixed configuration.
  • It does not sufficiently cover social fairness, impacts on people, accessibility or use of scores in high-risk decisions.
  • It does not resolve whether the object of measurement is an instance, a version, a family, a distribution of prompts or a population of executions.
  • It does not provide the data necessary to independently verify why each cell received its category.

Traceability

Scope: Full text

Version: INLG 2024 final proceedings paper, pp. 230–242, DOI 10.18653/v1/2024.inlg-main.19; 13 pages including two appendix tables

Consulted source: https://aclanthology.org/2024.inlg-main.19.pdf

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • No new model inference; conceptual framework and literature review
  • GPT-1, GPT-2, GPT-3, GPT-3.5, ChatGPT and GPT-4 across reviewed studies
  • BLOOM, FLAN-PaLM, PaLM, Chinchilla and Delphi across reviewed studies
  • BERT-family and distilled encoder models in the reviewed multilingual study

Instruments and metrics

  • Standards for Educational and Psychological Testing (AERA, APA and NCME, 2014)
  • International Guidelines for Test Use (International Test Commission, 2001)
  • R1 reliability for intended use
  • R2 validity for intended use
  • R3 suitability for test takers
  • R4 non-disclosure of test materials and contamination control
  • R5a test validity for all compared models
  • R5b validity of test translations
  • R5c transparent and reproducible test use
  • Five-level ordinal evidence rubric including not applicable

Data used

  • 25 machine-psychology papers retained from searches and citation tracing
  • 12 psychological constructs
  • 34 psychological tests and assessments
  • Table 1 study-by-requirement rating matrix
  • Appendix Table 2 construct-and-assessment map
  • Appendix Table 3 study, model and assessment map
  • No public machine-readable coding sheet or evidence-level artifact found

Evidence and location

  • Publication, authors, objective and main conclusion: INLG 2024 final paper, title page, abstract and section 1, pp. 230–231
  • Reference standards and distinction between reliability and validity: Final paper, section 2, p. 231
  • Requirements R1–R4: Final paper, sections 3.1–3.4, pp. 232–233
  • Requirements R5a–R5c: Final paper, section 3.5, pp. 233–234
  • Search, temporal cutoff, 25 papers, 12 constructs and 34 tests: Final paper, opening of section 4, p. 233
  • 20 studies with GPT-3 or later and 17 without another family: Final paper, section 4.1, pp. 234–235
  • Complete matrix and reconstructed counts by requirement: Final paper, Table 1, p. 234; visual transcription and independent count checked 15 Jul 2026
  • Joint assessment and semantics of the five categories: Final paper, section 4.2, p. 235
  • Findings on R1 and R2: Final paper, section 4.2, pp. 235–236
  • Findings on R3, R4, R5a, R5b and R5c: Final paper, section 4.2, p. 236
  • Nine modified tests, one validation, zero studies with all requirements and lenient criterion: Final paper, end of section 4.2 and opening of section 5, pp. 236–237
  • Open problems, non-exhaustive scope and conclusion: Final paper, sections 5–6, pp. 236–238
  • Map of 34 assessments and 25 studies/models: Final paper, Appendix Tables 2–3, pp. 241–242
  • Absence of associated public artifact: Exact-title, author, GitHub and OSF public search audited 15 Jul 2026; no official coding sheet, data or code located
  • Final bibliographic metadata: ACL Anthology 2024.inlg-main.19, pp. 230–242, DOI 10.18653/v1/2024.inlg-main.19, verified 15 Jul 2026