You don’t need a personality test to know these models are unreliable: Assessing the Reliability of Large Language Models on Psychometric Instruments

Evaluation and psychometric validity2024ACL AnthologyApproved editorial review

Original title: You Don't Need a Personality Test to Know These Models Are Unreliable: Assessing the Reliability of LLMs on Psychometric Instruments

Authors: Bangzhao Shu, Lechen Zhang, Minje Choi, Lavinia Dunagan, Lajanugen Logeswaran, Moontae Lee, Dallas Card, David Jurgens

Keywords: Large Language Models, Psychometric Instruments, Persona Measurement, Prompt Consistency, NLU

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

8
Authors
14
Findings
26
Limitations
15
Evidence

Editorial summary

English

Shu et al. ask a question that should precede any personality test applied to a language model: before interpreting a score, does the LLM understand the response format and preserve its answer when content is unchanged or reverse it when the proposition is reversed? They build MODEL-PERSONAS, 693 statements drawn from 39 instruments and organized into 115 axes. The collection is broader than personality: it combines traits, values, political and moral beliefs, social attitudes, descriptors, and situational responses. Items are standardized into binary Yes/No or True/False questions and tested through punctuation, whitespace, separator, answer-label, option-order, direct-negation, and paraphrastic-opposite variants.

The paper separates three properties. Comprehensibility is the proportion of responses whose first token is one of the allowed options; sensitivity is the fraction unchanged by spurious format variations; consistency is the fraction preserving an answer under equivalent labels or order and reversing it when meaning is reversed. This last quantity is within-item consistency, not alpha, omega, test–retest reliability, or cross-item internal consistency, a distinction the authors explicitly make. Seventeen models are tested at temperature 0. GPT-2 and RedPajama are removed from the main consistency analysis because of low valid-answer rates, so Figure 1 compares 15 models: 13 open models plus GPT-3.5 and GPT-4.

The results reveal substantial and practically important brittleness. BLOOMZ and FLAN-T5 usually emit valid option tokens, whereas Falcon, RedPajama, and several Llama 2 variants can move from valid answers to almost none after a space, line break, or terminal punctuation change. FLAN-T5 Base/Large/XL is the most robust family under spurious formatting. Option order and the True/False–Yes/No switch preserve many responses, although the paper correctly notes that an always-positive response policy can also score highly. Negation is the central failure: 10 of 15 models are near or below random behavior in at least these comparisons, and every model performs worse on paraphrastic reversal than on an explicit negation word.

The study manually selects 0.6 as a simultaneous threshold on four metrics and states that only FLAN-T5-XL, GPT-3.5, and GPT-4 pass, with FLAN-T5-Large close. This cutoff is neither calibrated nor preregistered and is not a psychometric criterion. Auditing the official artifact at commit fdfdf513a88de2af50294d10def9ca9ccfac87e8 reproduces the published processing logic but exposes boundary sensitivity: the released processed CSV yields about 0.605 paraphrastic-negation consistency for FLAN-T5-Large, whereas the released GPT-3.5 output yields about 0.592. Under a strict unrounded rule, membership differs from the prose. This does not overturn the general brittleness finding, but it does mean that the reported group of three is not a stable classification.

A second experiment prepends “You are a normal person,” one of six specific profiles, or a 35-attribute description. On average, every intervention lowers negation consistency; a few related axes improve as outliers, but the highly personified prompt is among the least consistent. The result supports a restrained conclusion: adding more personality adjectives does not itself create stable measurement. The experiments measure the robustness of prompt-conditioned binary answers, however; they do not show that a model lacks all behavioral tendencies, nor do they establish the presence or absence of a latent personality. The paper's reference contribution is therefore a set of format and polarity controls that should precede questionnaire interpretation, not an ontological validation of model personality.

Español

Shu et al. formulan una pregunta previa a cualquier test de personalidad aplicado a modelos: antes de interpretar una puntuación, ¿el LLM entiende el formato y mantiene la misma respuesta cuando el contenido no cambia o cuando la proposición se invierte? Para estudiarlo construyen MODEL-PERSONAS, 693 enunciados procedentes de 39 instrumentos y organizados en 115 ejes. El conjunto no se limita a personalidad: mezcla rasgos, valores, creencias políticas y morales, actitudes sociales, descriptores y respuestas situacionales. Los ítems se normalizan a una pregunta binaria Yes/No o True/False y se evalúan mediante cambios de puntuación, espacios, separadores, etiqueta de respuesta, orden de opciones y dos inversiones de significado: negación directa y paráfrasis de polaridad opuesta.

El paper separa tres propiedades. «Comprehensibility» es la proporción de respuestas cuyo primer token pertenece a las opciones permitidas; «sensitivity» es la fracción que no cambia ante variaciones espurias de formato; «consistency» es la fracción que conserva la respuesta ante etiquetas u orden equivalentes o la invierte cuando se invierte el significado. Esta última es consistencia intraítem, no alfa, omega, fiabilidad test–retest ni consistencia entre ítems; los autores lo explicitan. Se prueban 17 modelos a temperatura 0. GPT-2 y RedPajama se excluyen del análisis principal de consistencia por baja capacidad de producir respuestas válidas, de modo que la Figura 1 compara 15: 13 abiertos y GPT-3.5/GPT-4.

Los resultados muestran una fragilidad real y relevante. BLOOMZ y FLAN-T5 suelen devolver tokens válidos, mientras Falcon, RedPajama y varios Llama 2 pueden pasar de responder correctamente a casi no hacerlo por un espacio, salto de línea o signo final. FLAN-T5 Base/Large/XL es la familia más robusta a cambios espurios. El orden y el cambio True/False–Yes/No conservan muchas respuestas, aunque el propio artículo advierte que contestar siempre en la misma dirección también produce una consistencia alta. La prueba decisiva es la negación: 10 de 15 modelos quedan cerca o por debajo del azar en al menos estas comparaciones, y todos rinden peor con negación parafrástica que con una palabra negativa explícita.

El trabajo fija manualmente 0,6 como umbral simultáneo para cuatro métricas y afirma que solo FLAN-T5-XL, GPT-3.5 y GPT-4 lo cumplen, con FLAN-T5-Large cerca. Ese corte no está calibrado ni predefinido y no equivale a un criterio psicométrico. La auditoría del artefacto oficial, commit fdfdf513a88de2af50294d10def9ca9ccfac87e8, reproduce la lógica publicada pero encuentra casos frontera sensibles a datos/procesado: el CSV procesado liberado da aproximadamente 0,605 a FLAN-T5-Large en negación parafrástica, mientras el output liberado de GPT-3.5 da aproximadamente 0,592. Con una regla estricta sin redondear, la pertenencia al grupo cambia respecto a la prosa. Esto no invalida la fragilidad general, pero sí impide tratar el grupo de tres como resultado estable.

Una segunda prueba añade «You are a normal person», seis perfiles concretos o una descripción de 35 atributos. En promedio, todas esas intervenciones reducen la consistencia de negación; algunos ejes relacionados mejoran como valores atípicos, pero el prompt altamente personificado queda entre los peores. El resultado respalda una conclusión prudente: inyectar más adjetivos no crea por sí solo una medición estable. Sin embargo, las pruebas capturan robustez de respuestas binarias condicionadas por prompts, no demuestran que un modelo carezca de toda tendencia conductual ni que exista o no una personalidad latente. La aportación de referencia es, por tanto, el control de formato y polaridad que debería preceder a cualquier interpretación de cuestionarios; no una validación ontológica de la personalidad del modelo.

Research question

Do LLMs produce valid responses, robust to spurious format changes and coherent under equivalent or opposite semantic variations when administered personality instruments, and can a persona description in the prompt improve that consistency?

Method

Benchmark and perturbation study. The authors select 39 instruments of relatively stable attributes, exclude mental health, and normalize 693 items from 115 axes to binary prompts. They generate variants of punctuation, spaces, separators, Answer/Response label, option order, Yes/No versus True/False, direct negation, and paraphrastic reversal. Seventeen LLMs are run at temperature 0; valid initial token and proportion of responses that is preserved or inverted according to the transformation are measured. Fifteen models enter the main comparison of four consistencies. Then the change in negation consistency is calculated after prepending a normal, Big Five, conservative, or highly personified profile. The editorial review read and rendered the 19 pages, checked tables and appendices, downloaded the official repository and the external results archive, inspected the code, and recalculated the four consistencies from the released artifacts.

Sample: Table 1 covers 17 models: 15 open from six families plus GPT-3.5 and GPT-4. GPT-2 and RedPajama are excluded from the central analysis due to low comprehensibility, so Figure 1 contains 15 models. Each base unit comprises 693 items and five variants, 3,465 prompts; spurious formats are run separately. The persona section claims to use ten models, although the article does not unambiguously list that subset and the released external archive only retains results for five models for the published profiles. Inferences use temperature 0, Transformers 4.22.1, PyTorch 2.0.1, CUDA 11.7, and up to three RTX A6000; no seeds, exact weight revisions, or snapshots/dates of GPT-3.5 and GPT-4 are published.

Findings

  • BLOOMZ obtains comprehensibility 1.00 across all variants in Table 1; FLAN-T5 is also nearly perfect, except 0.76 for FLAN-T5-small with the double-bar separator.
  • Falcon-7B drops from 1.00 to 0.00 when introducing one or two spaces after the colon; RedPajama and GPT-2 generate very few valid tokens in numerous formats.
  • Llama 2 base is especially sensitive to the final punctuation mark: Llama2-7B goes from 1.00 with a colon to 0.00 with a question mark, and Llama2-13B from 1.00 to 0.03.
  • FLAN-T5 Base, Large, and XL preserve practically all of their responses under the majority of spurious changes; high comprehensibility does not necessarily imply low sensitivity in BLOOMZ.
  • The 15 models in Figure 1 exceed chance in order consistency and the majority exceed 0.7, but a constant response policy can also achieve this.
  • Consistency between True/False and Yes/No varies greatly: BLOOMZ-1B1 is high despite its low negation consistency, evidence of uniform responding rather than semantic understanding.
  • Ten of fifteen models fall near or below 0.5 in negation; only FLAN-T5-Large, FLAN-T5-XL, BLOOMZ-7B1, GPT-3.5, and GPT-4 stand out simultaneously in the two inversions according to the paper's description.
  • All models perform worse with paraphrastic reversal than with an explicit negation word.
  • The 0.6 threshold is chosen by manual inspection; the prose classifies FLAN-T5-XL, GPT-3.5, and GPT-4 as the only models that satisfy the four metrics.
  • Recalculation from the released artifact obtains approximately 0.605 for paraphrastic negation of FLAN-T5-Large and 0.592 for GPT-3.5; strict application of the cutoff changes the borderline cases relative to the prose.
  • Adding "You are a normal person", specific profiles, or 35 attributes reduces the overall distribution of negation consistency compared to the prompt without a persona.
  • Some axes related to the injected profile improve as outliers, for example extraversion under the extroverted profile, but the highly personified profile is among the least consistent.
  • No monotonic relationship appears between number of parameters and the four consistencies; results cluster better by architectural family.
  • The article itself recommends checking sensitivity and negation before interpreting personality, opinion, morality, or ideology scores obtained through prompts.

Limitations

  • MODEL-PERSONAS uses "persona" as an umbrella for traits, values, ideology, attitudes, gender, beliefs, and situations; a joint average does not represent a single psychological construct.
  • Converting originally Likert or multidimensional instruments to True/False eliminates intensity, scoring rules, reverse items, and validation properties of the original scales.
  • The consistency studied is intra-item under perturbation. It does not estimate internal consistency, test-retest reliability, longitudinal stability, inter-rater agreement, or measurement error of a trait score.
  • The title uses "reliability" in a broad sense; readers may confuse prompt robustness with psychometric reliability despite the correct distinction in section 4.3.
  • The same protocol is not administered to humans; the expectation that a person would respond identically to all paraphrases and negations is not empirically validated.
  • The 1,386 direct and paraphrastic inversions are described as manual, but no annotation authors, double coding, agreement, adjudication, or formal test of logical equivalence are reported.
  • The released paraphrases sometimes change intensity, focus, or pragmatics, not only polarity; for example substituting "important" with "trivial" or inverting comparisons may introduce a distinct semantic task.
  • A binary line does not capture ambivalent, conditional, or context-dependent responses; forcing a token can convert reasonable uncertainty into apparent inconsistency.
  • Temperature 0 and a single output per prompt offer no replications, variation across runs, temporal test-retest, or intervals from generation randomness.
  • The error bars in Figure 1 are not defined in the text or caption, and no contrasts, intervals, or correction for the numerous comparisons are reported.
  • The random baseline of 0.5 is descriptive; it does not model response bias, Yes/No prevalence, item difficulty, or dependence between questions.
  • The paper itself acknowledges that order and option may be high due to always answering positive; this response set is not corrected, nor is semantic information above a constant policy estimated.
  • The 0.6 cutoff was selected by manual inspection, without preregistration, psychometric justification, sensitivity to other cutoffs, or consistent treatment of rounding.
  • The statement that the majority does not exceed a random choice mixes four metrics: in order all exceed 0.5 and the main failure is concentrated in negation.
  • The persona test averages changes over heterogeneous models and instruments and presents boxplots without tests; the negative average may hide systematic improvements and deteriorations by axis.
  • The 35 attributes of the highly personified profile may be incompatible or semantically overlapping; its worst result does not separate length, instruction conflict, and number of traits.
  • The open models reach only 13B and belong to a 2022-2023 generation; generalization to larger current models, reasoners, or those with different chats/templates is limited.
  • Exact weight/tokenizer revisions, seeds, system prompts, chat templates, and concrete snapshots and dates of GPT-3.5/GPT-4 are not fixed.
  • The official code contains absolute paths from the author's environment and references data/templates.json, a file deleted before the audited commit; it also does not publish requirements or a lockfile, so it does not run from scratch without repair.
  • The inference script defines MAX_NEW_TOKENS=1 but does not use it, generates with total max_length 200, and keeps only the 50 tokens with the highest probability; these decisions are not documented in the paper with sufficient precision.
  • Postprocessing converts columns to lowercase and, when several capitalizations collide, takes the maximum instead of summing their mass; the results depend on a tokenization convention that is not analyzed.
  • The repository includes base results for 15 models, but not the code that produces tables, figures, error bars, thresholds, or persona tests from the CSVs.
  • The Google Drive archive called complete contains the base GPTs and part of the sensitivity analysis, but not all models/formats from Table 2 or the ten models declared for each persona profile.
  • Reproduction from released artifacts places two borderline cases on opposite sides of the cutoff relative to the prose, a sign of data/processing drift or undocumented rounding.
  • Permissions/licenses for republication of all items from the 39 instruments are not discussed, nor whether binarization preserves their conditions of use.
  • The results prove fragility under the chosen perturbations, not exhaustiveness: languages, conversational prompts, open responses, roles, few-shot, hidden chain-of-thought, and long contexts remain outside the scope.

What the study does not establish

  • It does not demonstrate that LLMs possess or lack conscious personality, identity, intentions, or human psychological states.
  • It does not demonstrate the absence of stable behavioral tendencies in tasks other than answering binary questionnaires.
  • It does not validate Big Five, MBTI, morality, ideology, or values scores for any of the evaluated models.
  • It does not test full psychometric reliability; it mainly measures intra-item robustness to format and semantic reversal.
  • It does not demonstrate that 0.6 is a scientifically valid threshold or that models situated above it have a consistent persona.
  • It does not allow the claim that FLAN-T5-XL, GPT-3.5, and GPT-4 form a stable group of passing models: the released artifacts change the borderline cases.
  • It does not demonstrate that a response that inverts under negation reflects deep understanding; it may exploit the lexical cue, especially in direct negation.
  • It does not demonstrate that a response that does not invert is always erroneous, because some paraphrases alter strength, scope, or pragmatics.
  • It does not causally prove that architecture explains the familial differences; training, instructions, tokenizer, chat tuning, and scale are confounded.
  • It does not prove that adding a persona always worsens all axes; it reports a mean distribution with related exceptions.
  • It does not generalize directly to later models, other languages, open responses, tools, or longitudinal interactions.
  • It does not allow exact reproduction of all tables and figures from the repository and the external archive without additional decisions and materials.
  • It does not justify substituting human responses with LLM outputs in surveys, public policy, or decisions about social groups.
  • It does not by itself invalidate all use of human instruments with LLMs; it establishes minimum controls that any serious use must surpass and document.

Traceability

Scope: Full text

Version: NAACL 2024 long paper, pp. 5263–5281, DOI 10.18653/v1/2024.naacl-long.295; final 19-page proceedings version with appendices

Consulted source: https://aclanthology.org/2024.naacl-long.295.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

  • GPT-2
  • tiiuae/falcon-7b-instruct
  • togethercomputer/RedPajama-INCITE-7B-Instruct
  • bigscience/bloomz-560m
  • bigscience/bloomz-1b1
  • bigscience/bloomz-3b
  • bigscience/bloomz-7b1
  • Llama 2 7B
  • Llama 2 7B Chat
  • Llama 2 13B
  • Llama 2 13B Chat
  • google/flan-t5-small
  • google/flan-t5-base
  • google/flan-t5-large
  • google/flan-t5-xl
  • OpenAI GPT-3.5
  • OpenAI GPT-4

Instruments and metrics

  • MODEL-PERSONAS: 39 source instruments, 115 persona axes and 693 standardized statements
  • Five construct groups: Beliefs, Normative statements, Values, Descriptors and Situations
  • Binary Yes/No and True/False response formats
  • Comprehensibility as valid first-option token rate
  • Sensitivity under punctuation, whitespace, separator and response-label changes
  • Order consistency
  • Option-label consistency
  • Direct-negation consistency
  • Paraphrastic-negation consistency
  • Manually selected four-metric threshold of 0.6
  • Persona-prompt consistency shift

Data used

  • 693 standardized English statements from 39 psychological and sociological instruments
  • 115 axes spanning traits, values, beliefs, norms, attitudes and situations
  • 3,465 inputs per base inference unit: 693 statements × 5 content/option variants
  • Nine additional spurious prompt-format variants
  • Direct and paraphrastic opposite versions for all 693 statements
  • Official llm-personas repository at commit fdfdf513a88de2af50294d10def9ca9ccfac87e8
  • Official Google Drive result archive, SHA-256 5c009cbedffd6f0ce01bf33fbba6af328b3194b564e7c1d210b05bbbbad40795

Evidence and location

  • Publication, authors, DOI, 17 models, and objective: NAACL 2024 final paper, title page and abstract, p. 5263
  • 39 instruments, 693 questions, 115 axes, and five groups: Final paper, section 2, pp. 5264-5265; Appendix Table 3, p. 5277
  • Base prompt, prior elimination of negations, and spurious variants: Final paper, sections 2-3.1, p. 5265; Appendix Table 4, p. 5277
  • Order, options, direct and paraphrastic negation: Final paper, section 3.2, pp. 5265-5266; Appendix Table 5, p. 5277
  • Definitions of comprehensibility, sensitivity, and intra-item consistency: Final paper, sections 4.1-4.3, p. 5266
  • Models, temperature, and execution details: Final paper, section 4.4, pp. 5266-5267; Appendix E, pp. 5275-5276
  • Comprehensibility and sensitivity by format: Final paper, Tables 1-2 and sections 5.1-5.2, pp. 5267-5268
  • Four consistencies, negation, and random baseline: Final paper, Figure 1 and section 5.3, pp. 5268-5269
  • Manual 0.6 threshold and classification of three models: Final paper, end of section 5.3, p. 5269
  • Persona profiles and general decrease in consistency: Final paper, section 6 and Figure 2, pp. 5269-5270; Appendix Table 6, p. 5278
  • Size, family, conclusions, limitations, and ethical considerations: Final paper, sections 7-10, pp. 5270-5272; Appendix Figures 3-6, pp. 5278-5279
  • Variation by axis: Final paper, Appendix Figures 7-8, pp. 5280-5281
  • Reproducibility, code, and official data: Official orange0629/llm-personas repository commit fdfdf513a88de2af50294d10def9ca9ccfac87e8 and linked Google Drive archive, audited 15 Jul 2026
  • Recalculated threshold borderline cases: Released all_truefalse_short_changeline.csv and GPT-3.5 raw CSV joined to the 3,465-row official prompt manifest; paraphrastic consistency approximately 0.605 for FLAN-T5-Large and 0.592 for GPT-3.5
  • Final bibliographic metadata: ACL Anthology 2024.naacl-long.295, DOI 10.18653/v1/2024.naacl-long.295, verified 15 Jul 2026