A Survey of Large Language Models for Perception and Measurement of Human Psychology

Reviews, theory, and governance2026arXivApproved editorial review

Authors: Yudong Li, Xiaoyi Chen, Jiawei Cai, Zehao Zhong, Haoyang Yang, Huajin Tang, Linlin Shen

Keywords: Psychological measurement, Personality assessment, Mental health screening, Psychometrics, Narrative survey, Review methodology

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

7
Authors
12
Findings
15
Limitations
7
Evidence

Editorial summary

English

This 21-page paper, accepted by IEEE Transactions on Cognitive and Developmental Systems and available as arXiv v1, organizes literature on language models used as instruments to perceive or measure human psychology. It asks whether an LLM can infer latent constructs, personality, emotion, cognitive states, or mental-health indicators, from conversation, natural language, or multimodal signals. The authors structure the field around theoretical plausibility (why measurement might work), measurement methodology (how it is performed), and application effectiveness (what has been measured). The first invokes functional theory-of-mind results and virtual-subject simulation; the second distinguishes active conversational assessment, passive natural-language assessment, and multimodal fusion; the third covers personality and mental health. The source cites 229 unique references and includes six tables: theory-of-mind results, virtual subjects, 24 studies organized by paradigm, nine personality frameworks, twenty “performance trend” rows, and fifteen real or synthetic datasets.

The active-assessment synthesis covers adaptive interviews and dialogues, chain-of-thought prompting, instruction tuning, and multi-agent systems that ask, clarify, and score. Passive assessment includes zero/few-shot classification, embeddings, RAG, fine-tuning, and distillation to infer traits or symptoms from essays, social media, and transcripts. The multimodal section combines text with images, video, audio, wearables, or physiological signals. The personality discussion covers the Big Five, HEXACO, MBTI, Dark Triad, and less frequently used frameworks; the mental-health section includes depression, anxiety, PTSD, emotion, and suicide-risk detection. The paper acknowledges major limitations: prompt sensitivity, temporal instability, socially desirable responding, reverse-coded-item failures, hallucination, opacity, privacy, cultural and demographic bias, cost, vendor dependence, and insufficient clinical validation. Its final position is appropriately cautious: under controlled conditions LLMs may support structured inference or screening tasks, but they do not meet the reliability and interpretability requirements for replacing validated instruments or clinical judgment.

The audit changes how the contribution should be characterized. Although the paper repeatedly calls itself a “systematic review,” it reports no databases, search strings, search dates, inclusion or exclusion criteria, deduplication, screening stages, reviewer counts, disagreement resolution, flow diagram, registered protocol, extraction form, study-quality assessment, or risk-of-bias method. It also releases no complete inventory of selected studies or extraction dataset. It is therefore a broad narrative survey with a useful taxonomy, not a reproducible systematic review. Readers cannot determine whether coverage is exhaustive, what was excluded, or how selection bias affects the conclusions.

The review also mixes distinct research objects. Its declared focus is using LLMs to measure people, but the theoretical basis and several applications include models acting as virtual subjects, questionnaires administered to the LLM itself, personality role-play, and multi-agent social simulation. A model's ability to produce persona-consistent answers or solve a theory-of-mind benchmark does not validate measurement of a real human's personality or mental health. Parts of the multimodal section cover sentiment analysis about products, events, or celebrities, while the mental-health history includes BERT and RoBERTa systems. These are relevant context but extend beyond LLM-based human psychometrics and cannot be pooled as equivalent validation evidence.

The psychometric evidence does not support a uniform quantitative conclusion. The framework table assigns High, Moderate-High, Moderate, or Low validation labels without a reported rubric. The performance table mixes accuracy, MSE, F1, percentage improvement, narrative descriptions, trait scores, agreement, R-squared, correlations, Cronbach's alpha, explained variance, and Cohen's d across incompatible datasets, tasks, and model classes, including pre-LLM methods. It therefore cannot demonstrate steady improvement over time or rank systems. In mental health, the text describes partial concurrent validity because some classifiers match supervised baselines, but later acknowledges that sensitivity, specificity at clinically meaningful thresholds, predictive validity, and cross-model agreement remain underexplored. Benchmark accuracy is not diagnostic validity, calibration, clinical benefit, or safe screening performance. “Moderate validity” is a narrative judgment, not a pooled effect or formal certainty grade.

Several quantitative and regulatory claims need explicit boundaries. The cost example says 10,000 transcripts averaging 500 input tokens would cost about $375 using GPT-4. At the paper's own stated rate of $30 per million input tokens, five million input tokens cost $150; output-token quantity and price are omitted, so $375 is not reproducible. In the United States, the survey broadly states that FDA classifies clinical decision-support software intended to inform diagnosis or treatment as a device, but official guidance distinguishes functions meeting all four statutory Non-Device CDS criteria from device software functions. In the EU, a health-related context alone is not a universal high-risk test: Article 6 relies on Annex I product/safety-component and conformity-assessment conditions or Annex III uses, with stated exceptions. The cited 2013 APA telepsychology guideline does support preserving reliability, validity, and administration conditions when tests are adapted to technology, but it predates LLM assessment and does not by itself support the broader automated-tool/clinical-judgment statement. Finally, the absolute assertion that no LLM psychological-assessment tool has regulatory clearance is not backed by a jurisdiction, registry, product definition, search date, or reproducible search.

The defensible contribution is a wide conceptual map and a well-directed warning about current limitations. It is not a new psychometric trial, meta-analysis, demonstration of clinical validity, or proof of exhaustive literature coverage, and it does not support replacing human assessment. Its strongest use is as a thematic entry point and provisional taxonomy, with every consequential claim checked against the cited primary study.

Español

Este trabajo, aceptado por IEEE Transactions on Cognitive and Developmental Systems y disponible como arXiv v1 de 21 páginas, organiza la literatura sobre modelos de lenguaje como instrumentos para percibir o medir atributos psicológicos humanos. Su pregunta es si un LLM puede inferir constructos latentes, personalidad, emoción, estados cognitivos o indicadores de salud mental, a partir de conversación, texto natural o señales multimodales. Los autores estructuran el campo en tres dimensiones: plausibilidad teórica (por qué podría funcionar), metodología de medición (cómo se realiza) y eficacia aplicada (qué se ha medido). La primera recurre a resultados funcionales de teoría de la mente y a la capacidad de simular “sujetos virtuales”; la segunda distingue evaluación activa y conversacional, evaluación pasiva de lenguaje y fusión multimodal; la tercera recorre personalidad y salud mental. La fuente cita 229 referencias únicas y contiene seis tablas: una de teoría de la mente, otra de sujetos virtuales, 24 estudios organizados por paradigma, nueve marcos de personalidad, veinte filas de “tendencias” de rendimiento y quince datasets reales o sintéticos.

La síntesis describe, dentro de la evaluación activa, entrevistas o diálogos adaptativos, prompting con chain-of-thought, ajuste de instrucciones y arquitecturas multiagente que preguntan, aclaran y puntúan. En evaluación pasiva reúne clasificación zero/few-shot, embeddings, RAG, ajuste fino y destilación para inferir rasgos o síntomas desde ensayos, redes sociales y transcripciones. La sección multimodal combina texto con imagen, vídeo, audio, wearables o señales fisiológicas. Para personalidad, repasa Big Five, HEXACO, MBTI, Dark Triad y marcos menos utilizados; para salud mental, incluye detección de depresión, ansiedad, PTSD, emoción y riesgo suicida. El artículo reconoce problemas sustantivos: sensibilidad al prompt, inestabilidad temporal, respuestas socialmente deseables, dificultad con ítems invertidos, alucinación, opacidad, privacidad, sesgo cultural y demográfico, coste, dependencia de proveedores y falta de validación clínica. Su posición final es prudente: bajo condiciones controladas los LLM pueden apoyar tareas estructuradas de inferencia o cribado, pero no alcanzan la fiabilidad e interpretabilidad necesarias para sustituir instrumentos validados o juicio clínico.

La auditoría obliga, sin embargo, a cambiar cómo debe describirse la contribución. Aunque el artículo se llama repetidamente “systematic review”, no informa bases de datos, cadenas de búsqueda, fechas, criterios de inclusión o exclusión, deduplicación, fases de cribado, número de revisores, resolución de desacuerdos, diagrama de flujo, protocolo registrado, formulario de extracción, evaluación de calidad ni riesgo de sesgo. Tampoco publica el inventario completo de estudios seleccionados ni un dataset de extracción. Es por tanto una encuesta narrativa amplia con una taxonomía útil, no una revisión sistemática reproducible. No se puede saber si la cobertura es exhaustiva, qué trabajos fueron descartados o cuánto sesgo de selección afecta a las conclusiones.

También hay mezcla de objetos de estudio. El foco declarado es usar LLM para medir a personas, pero la base teórica y varias aplicaciones incluyen modelos como sujetos virtuales, cuestionarios administrados al propio LLM, role-play de personalidades y simulaciones multiagente. Que un modelo reproduzca respuestas coherentes con una persona o resuelva una prueba de teoría de la mente no demuestra que mida válidamente la personalidad o salud mental de un humano real. Parte de la sección multimodal incluye sentiment analysis sobre productos, eventos o celebridades, y la historia de salud mental incorpora BERT/RoBERTa; eso amplía el alcance más allá de LLM y psicometría humana. Estas líneas son relevantes como contexto, pero no deben agregarse como evidencia equivalente de validez de medida.

La evidencia psicométrica tampoco permite una conclusión cuantitativa uniforme. La tabla de marcos asigna etiquetas High, Moderate-High, Moderate o Low sin rubricar cómo se obtuvieron. La tabla de rendimiento mezcla accuracy, MSE, F1, mejora porcentual, descripciones narrativas, scores de rasgo, acuerdo, R², correlaciones, alfa de Cronbach, varianza explicada y d de Cohen en datasets, tareas y modelos incompatibles, incluidos métodos anteriores a los LLM. Por ello no demuestra una mejora sostenida en el tiempo ni permite ordenar sistemas. En salud mental, el texto habla de validez concurrente parcial porque algunos clasificadores igualan baselines supervisados, pero luego admite que sensibilidad, especificidad en umbrales clínicos, validez predictiva y acuerdo entre modelos están poco estudiados. Accuracy de benchmark no equivale a diagnóstico válido, calibración, beneficio clínico o cribado seguro. La frase “validez moderada” es una valoración narrativa, no un efecto combinado ni una graduación formal de certeza.

Hay además afirmaciones cuantitativas y regulatorias que requieren límites. El ejemplo de coste dice que 10.000 transcripciones de 500 tokens de entrada costarían unos 375 dólares con GPT-4; con el precio que el propio texto cita, 30 dólares por millón de tokens de entrada, cinco millones de tokens son 150 dólares. Como no especifica tokens ni precio de salida, la cifra no se reproduce. En regulación estadounidense, el artículo afirma de forma general que la FDA clasifica como dispositivo el software de apoyo clínico destinado a informar diagnóstico o tratamiento, pero la guía oficial distingue funciones Non-Device CDS que cumplen los cuatro criterios legales de otras funciones que sí son dispositivo. En la UE, “usarse en un contexto sanitario” no basta por sí solo: el artículo 6 de la Ley de IA remite a condiciones de producto/componente de seguridad y evaluación de conformidad del Anexo I o a usos del Anexo III, con excepciones. La guía APA de telepsicología de 2013 sí exige preservar fiabilidad, validez y condiciones de administración al adaptar tests a tecnología, pero no trata LLM ni fundamenta por sí sola la frase más amplia sobre herramientas automatizadas como mero suplemento del juicio clínico. Finalmente, la afirmación absoluta de que ningún evaluador psicológico basado en LLM tiene autorización regulatoria no incluye jurisdicción, registro, definición de producto, fecha ni búsqueda verificable.

La contribución defendible es un mapa conceptual amplio y una advertencia bien orientada sobre los límites actuales. No aporta un nuevo ensayo psicométrico, no hace meta-análisis, no demuestra validez clínica ni exhaustividad de la literatura y no justifica sustituir evaluaciones humanas. Su mejor uso es como puerta de entrada temática y taxonomía provisional, verificando cada estudio primario y tratando sus conclusiones agregadas como narrativas.

Research question

What theoretical plausibility, methodological paradigms, applications, and limits does the literature that uses LLMs as instruments to infer or measure human psychological attributes present?

Method

Narrative survey organized into theoretical plausibility, measurement methodology, and applied efficacy. It classifies methods into active/conversational evaluation, passive language analysis, and multimodal fusion, and reviews personality, mental health, and technical, ethical, and regulatory challenges. The audit visually read the 21 pages, inspected all the TeX, the six tables, and 229 cited references, verified the absence of a systematic protocol, and contrasted regulatory claims with FDA, EUR-Lex, and APA.

Sample: There is no primary sample or meta-analysis. The article synthesizes 229 cited references in narrative form and selects examples for six tables; it does not publish the searched universe, the number screened, the excluded ones, or a reproducible inventory of included studies.

Findings

  • The three-dimensional and three-paradigm taxonomy offers a useful map of the field.
  • The authors prudently conclude that LLMs can only complement structured tasks under controlled conditions and cannot replace validated instruments or clinical judgment.
  • The label "systematic review" is not backed by a reported method of search, selection, extraction, or bias assessment.
  • The review mixes measurement of humans, simulation of subjects, and evaluation of the LLM's own psychology.
  • Functional theory of mind and role-play do not demonstrate construct validity for measuring real persons.
  • The trends table mixes incompatible metrics and tasks and does not demonstrate sustained improvement.
  • The qualitative labels for framework validation lack a published rubric.
  • The mental health evidence relies mostly on classification performance, not on complete clinical validation.
  • The cost of 375 dollars is not reproduced with the printed tokens and entry price.
  • The formulations on the FDA and the European AI Law are broader than the official rules.
  • The APA citation does support preserving psychometric properties when adapting tests to technology, but not a specific rule on LLMs.
  • No code, extraction, protocol, or analysis is published that would allow reproducing the coverage or synthesis.

Limitations

  • No databases, searches, dates, or documented time window.
  • No inclusion/exclusion criteria or flow diagram.
  • No double review, disagreement resolution, or registered protocol.
  • No quality assessment or risk of bias of primary studies.
  • No complete extraction table linked to each claim.
  • No meta-analysis or comparable measures across studies.
  • Inconsistent scope among measuring humans, simulating humans, and measuring models.
  • Inclusion of general sentiment analysis and non-LLM models.
  • Psychometric validation labels without operational criterion.
  • Performance trends constructed with heterogeneous metrics.
  • Clinical validity inferred partially from benchmark accuracy.
  • Cost estimates without sufficient assumptions and with non-reproducible arithmetic.
  • Absolute or overly broad regulatory claims.
  • No registry search supporting the global absence of authorizations.
  • No code, review data, or analytical notebook.

What the study does not establish

  • That the review is systematic or exhaustive.
  • Absence of bibliographic selection bias.
  • An aggregate estimate of validity or reliability.
  • Sustained improvement of the field between 2015 and 2025.
  • That theory of mind in LLMs validates human psychological measurement.
  • Equivalence between virtual subjects and human participants.
  • Diagnostic validity, calibration, or clinical utility.
  • Safety for screening, diagnosis, or autonomous intervention.
  • Cross-cultural validity or group invariance.
  • That any CDS is a medical device in the U.S.
  • That any AI in health is high-risk in the EU.
  • Global and current absence of any regulatory authorization.
  • Correctness of the published cost figures.
  • Suitability to replace psychometric instruments or professional judgment.

Traceability

Scope: Full text

Version: arXiv:2606.20603v1, 21 pages; accepted by IEEE TCDS; DOI 10.1109/TCDS.2026.3695985; complete TeX and all tables audited

Consulted source: https://arxiv.org/abs/2606.20603

Review: Codex 21-page visual, complete TeX/table, review-method, psychometric-claim, cost-arithmetic, FDA, EU AI Act, APA and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5 and GPT-4-family models cited across primary studies
  • Gemini-family models cited across primary studies
  • Llama-family models cited across primary studies
  • DeepSeek-family models cited across primary studies
  • Domain-tuned psychology and mental-health language models
  • BERT and RoBERTa systems as historical or comparative non-generative baselines

Instruments and metrics

  • Theory-of-mind benchmarks as theoretical context
  • Big Five
  • HEXACO
  • MBTI
  • Dark Triad / Short Dark Triad
  • Depression, anxiety and PTSD classification tasks
  • Emotion and empathy benchmarks
  • Accuracy, F1, MSE, R-squared, correlation and reliability metrics across heterogeneous studies
  • Independent systematic-review-method and regulatory-claim audit

Data used

  • No primary experimental dataset introduced by the survey
  • Twenty-four study rows in the paradigm table
  • Fifteen real or synthetic datasets in the personality/mental-health dataset table
  • Examples include Essays, CPED, PsyQA, PANDORA, CMACD, PDCH, EATD, myPersonality, MMPsy, SoulChat and Psych8k

Evidence and location

  • Text, structure, tables, conclusions, limitations, and references: arXiv:2606.20603v1, 21 pages, sha256 284ee1f7f2123022a427aa96401885901b26bcaf2a5babbc59b7c6d8abd4aa48
  • TeX source, absence of systematic method, and composition of tables: arXiv source v1 sha256 45514c47f863cd9a2e3fcad5c48d28888fc573594803cdbcab4be9238aa5cb59; main TeX sha256 f9d0588337db59c2654fd2bf1a27b8dcd821ad4eef53ca880fcd1f89ee9236eb
  • Acceptance and editorial metadata: arXiv record and DOI 10.1109/TCDS.2026.3695985, IEEE record 11534094, checked 2026-07-17
  • Regulatory scope of CDS in the United States: FDA Clinical Decision Support Software guidance, January 2026, checked 2026-07-17
  • High-risk classification rules in the EU: Regulation (EU) 2024/1689, Article 6 and Annexes I/III, official EUR-Lex text, checked 2026-07-17
  • Psychometric integrity when adapting assessment to telepsychology: APA Guidelines for the Practice of Telepsychology, American Psychologist 68(9), 2013, Guideline 6
  • Complete independent audit: reports/verification/article-323-llm-psychology-survey-method-scope-psychometrics-regulatory-and-reproducibility-audit.json