Synthetic Personality Review

Self-Assessment Tests are Unreliable Measures of LLM Personality

Evaluation and psychometric validity2024ACL AnthologyApproved editorial review

Original title: Self-Reports are Unreliable Measures of LLM Personality

Authors: Akshat Gupta, Xiaoyang Song, Gopala Anumanchipalli

Keywords: Large Language Models, Personality Assessment, Prompt Sensitivity, Option-Order Symmetry, NLP

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

3
Authors
13
Findings
23
Limitations
12
Evidence

Editorial summary

English

Gupta, Song, and Anumanchipalli subject direct administration of human questionnaires to two simple controls: prompt sensitivity and symmetry under option order. They use all 300 IPIP-NEO items, 60 for each OCEAN domain, with ChatGPT/gpt-3.5-turbo and Llama 2 Chat at 7B, 13B, and 70B. Every item is presented through three templates taken from prior work: choose A–E for the accuracy of a self-description, answer 1–5 for how much a described person resembles the model, and rate 1–5 agreement with the statement. They then reverse A–E order, the meaning of the 1–5 anchors, or the agree–disagree direction, and recode scores so that 5 continues to represent greater trait presence.

Mean scores change materially. For ChatGPT, Openness moves from 4.48 under Prompt 1 to 3.32 under Prompt 2 and 2.57 under Prompt 3; Conscientiousness moves from 4.35 to 3.30 and 2.53; Extraversion from 4.57 to 3.22 and 2.47. Reversal also shifts results: original Prompt 3 yields 2.47–2.68 across ChatGPT traits, while its reversed form yields 3.22–3.60. The Llama models show different patterns rather than a common scale. Using Mann–Whitney U at alpha 0.05, the paper reports differences in 29/30 contrasts for ChatGPT, 24/30 for Llama-2-7B, 26/30 for 13B, and 19/30 for 70B. This strongly supports the claim that an isolated score depends on administrative choices that should not define a personality.

The published statistical inference does not match the design, however. Each comparison contains the same items under two conditions, so observations are paired; Mann–Whitney assumes independent samples. The analysis also leaves 30 tests per model uncorrected, reports no effect sizes or intervals, and silently drops invalid outputs. The prose says each distribution contains 60 observations, but the official artifact yields only 6,891 scorable responses out of 7,200: ChatGPT retains 1,799/1,800 and the three Llama models retain 1,673, 1,708, and 1,711; some cells contain 47–59 responses rather than 60.

The audit recalculated all contrasts from the same outputs while pairing by item. Uncorrected Wilcoxon tests retain 29/30 differences for ChatGPT, 23/30 for 7B, 24/30 for 13B, and 21/30 for 70B. After Benjamini–Hochberg correction within each model's 30 contrasts, 29, 22, 23, and 18 remain. The main pattern survives, but the published counts should not be repeated as an exact reliability result. The three templates are also not strictly equivalent: they ask about descriptive accuracy, resemblance to another person, and agreement, with different anchors and pragmatics. Some variation may be undesirable prompt sensitivity, while some may be a response to genuinely different measurement frames.

The paper is right to require personality scores to survive reasonable controls and to note that items such as liking to be the center of attention presuppose autobiographical memory and introspection. It nevertheless studies one inventory, one language, four 2023-era models, six formats, and one near-deterministic run. It establishes that this direct administration of IPIP-300 does not yield robust scores under the tested conditions; it does not show that every future method for measuring model tendencies is impossible, nor does it resolve what machine personality should mean.

Español

Gupta, Song y Anumanchipalli someten la administración directa de cuestionarios humanos a dos controles sencillos: sensibilidad al prompt y simetría ante el orden de las opciones. Usan los 300 ítems IPIP-NEO, 60 por cada dominio OCEAN, con ChatGPT/gpt-3.5-turbo y Llama 2 Chat de 7B, 13B y 70B. Cada ítem se presenta mediante tres plantillas tomadas de trabajos previos: elegir A–E según la exactitud de una autodescripción, responder 1–5 según cuánto se parece al modelo una descripción y puntuar 1–5 el grado de acuerdo. Después se invierte el orden de A–E, el significado de 1–5 o la dirección agree–disagree, y el scoring se vuelve a orientar para que 5 conserve el significado de mayor presencia del rasgo.

Las medias cambian de forma material. En ChatGPT, por ejemplo, la apertura pasa de 4,48 con Prompt-1 a 3,32 con Prompt-2 y 2,57 con Prompt-3; consciencia de 4,35 a 3,30 y 2,53; extraversión de 4,57 a 3,22 y 2,47. La inversión también desplaza resultados: el Prompt-3 original de ChatGPT produce 2,47–2,68 según rasgo y el invertido 3,22–3,60. Los Llama muestran otros patrones, no una escala común. Con Mann–Whitney U a α=0,05, el paper declara diferencias en 29/30 contrastes para ChatGPT, 24/30 para Llama-2-7B, 26/30 para 13B y 19/30 para 70B. Esta evidencia apoya con fuerza que un score aislado depende de decisiones administrativas que no deberían definir una personalidad.

La inferencia estadística publicada, sin embargo, no respeta el diseño. Cada comparación contiene los mismos ítems bajo dos condiciones, por lo que las observaciones están emparejadas; Mann–Whitney supone muestras independientes. Tampoco se corrigen 30 contrastes por modelo, no se informan tamaños de efecto o intervalos, y los outputs inválidos se eliminan silenciosamente. El texto dice que cada distribución contiene 60 observaciones, pero el artefacto oficial solo permite puntuar 6.891 de 7.200 respuestas: ChatGPT conserva 1.799/1.800 y los Llama 1.673, 1.708 y 1.711; algunas celdas tienen 47–59 respuestas, no 60.

La auditoría recalculó los contrastes en los mismos outputs, emparejando por ítem. Wilcoxon sin corrección mantiene 29/30 diferencias en ChatGPT, 23/30 en 7B, 24/30 en 13B y 21/30 en 70B. Tras Benjamini–Hochberg dentro de los 30 contrastes de cada modelo quedan 29, 22, 23 y 18. El patrón principal sobrevive, pero las cifras publicadas no deben repetirse como si fueran una prueba exacta de fiabilidad. Además, las tres plantillas no son estrictamente equivalentes: preguntan por exactitud descriptiva, semejanza con una persona y acuerdo, con anclajes y pragmática distintos. Parte de la variación puede ser sensibilidad indeseable y parte respuesta a marcos de medida realmente diferentes.

El trabajo acierta al exigir que cualquier score de personalidad sea estable ante controles razonables y al recordar que preguntas como «me gusta ser el centro de atención» presuponen memoria autobiográfica e introspección. No obstante, estudia una sola escala, un idioma, cuatro modelos de 2023, seis formatos y una ejecución casi determinista. Demuestra que esta administración directa de IPIP-300 no ofrece scores robustos bajo sus condiciones; no demuestra que toda forma futura de medir tendencias de un modelo sea imposible ni resuelve qué significaría «personalidad» en una máquina.

Research question

Do the Big Five scores obtained by administering IPIP-300 directly to an LLM remain comparable when switching among three plausible templates and when reversing the order or direction of their response options?

Method

Factorial robustness experiment of administration. Four chat models respond to the 300 IPIP-NEO items under three original prompts and their three reversed order/scale versions, for a nominal maximum of 7,200 responses. Inverse items and inverted scales are recoded to 1–5; mean and deviation are calculated per 60 items of each trait. The article uses Mann–Whitney U for three template comparisons and three original–reversed comparisons per trait and model. The audit read and rendered the 14 pages, examined the official repository at commit f5da8dc08f890c8c6df1de57e0cd55143b89768d, reproduced Table 3, counted scorable outputs and repeated the 30 contrasts of each model with paired Wilcoxon and Benjamini–Hochberg correction.

Sample: There are no human participants. The repeated unit is each of the 300 fixed items, administered six times to each model; each trait contributes at most 60 pairs. ChatGPT retains 1,799 of 1,800 responses and Llama-2-7B, 13B and 70B retain 1,673, 1,708 and 1,711, respectively. Per condition/trait, ChatGPT has 59–60 scores, Llama-2-7B 47–60, 13B 50–60 and 70B 51–60. For the paired reanalysis, the intersections of valid responses span 43–60 items. The paper declares temperature 0.01 and top-p 1; the Llama code passes those values, but the OpenAI script does not pass the variable and executes the function's default value, temperature 0.

Findings

  • ChatGPT systematically shifts from high scores with Prompt-1 to intermediate with Prompt-2 and low with Prompt-3: O 4.48/3.32/2.57; C 4.35/3.30/2.53; E 4.57/3.22/2.47; A 3.72/3.08/2.68; N 4.27/3.20/2.52.
  • The inverted scale of Prompt-3 shifts ChatGPT to O 3.60, C 3.53, E 3.60, A 3.22 and N 3.55 even after recoding the direction.
  • Llama-2-13B shows an extreme contrast between Prompt-2 and Prompt-3: openness 2.11 versus 4.43 and neuroticism 2.47 versus 4.64.
  • The three Llama sizes do not follow a common monotonic pattern; each template and order interacts differently with each model.
  • The paper reproduces 15/15 significant sensitivity contrasts in ChatGPT, 11/15 in Llama-2-70B, 15/15 in 13B and 11/15 in 7B.
  • For original versus reversed order/scale, Mann–Whitney marks 14/15 in ChatGPT, 8/15 in 70B, 11/15 in 13B and 13/15 in 7B.
  • The published totals are therefore 29/30, 19/30, 26/30 and 24/30 for ChatGPT, Llama-2-70B, 13B and 7B, respectively.
  • Those same totals are reproduced from the released JSON and parsers, although the statistical and figure scripts are not published.
  • Paired Wilcoxon without correction gives 29/30, 21/30, 24/30 and 23/30 in that same model order.
  • After Benjamini–Hochberg on the 30 tests of each model, 29/30, 18/30, 23/30 and 22/30 remain; the dependence on format remains extensive.
  • The parsers discard 309 of 7,200 outputs and the discard rate depends on model and prompt, so the ability to obey the format is also part of the result.
  • The absence of ground truth prevents selecting one prompt as the correct version just because it produces more intuitive scores; stability across administrations is a reasonable prerequisite.
  • The article concludes that the personality of these LLMs should not be quantified with isolated IPIP self-report scores and calls for more robust measures.

Limitations

  • The conditions are matched by item, but Mann–Whitney U treats the samples as independent; Wilcoxon or a paired permutation test better respects the design.
  • Mann–Whitney is not in general a test of full equality of distributions as the null hypothesis of the text claims; without additional assumptions it evaluates stochastic order/ranks.
  • 30 tests per model and 120 in total are performed at α=0.05 without correction for multiplicity.
  • No paired effect sizes, intervals, mean absolute differences, correlations between administrations or a prior criterion of practical equivalence are reported.
  • The 60 items are a fixed set from the instrument, not 60 independent subjects; the goal of statistical generalization to a universe of items is not defined.
  • The text claims that each distribution has 60 samples, but the parsers remove responses that do not contain exactly one recognizable option and leave up to 47 per cell.
  • The removal depends on model and prompt and is not analyzed as missingness; comparing only valid responses can change means and p-values in a non-random way.
  • The three templates are not semantically identical: descriptive accuracy, likeness to a person and agreement activate different pragmatic frames and anchors.
  • Prompt-1 and Prompt-2 also differ in item position, indices, separators, uncertainty instructions and neutral category wording; the experiment does not isolate which component causes the change.
  • The authors adapt prompts taken from different studies to five points, so they do not exactly replicate all the original administrations that are cited.
  • Means and deviations on ordinal scores are descriptive, but do not model thresholds, acquiescence, category use or differential item functioning.
  • A nearly deterministic run per condition does not measure temporal stability, across servers, seeds or samplings; the paper itself acknowledges that higher temperatures change responses.
  • The paper declares temperature 0.01 for all, but query_openai_api.py calls ask_gpt_chat without passing it and uses the default 0; the actual ChatGPT configuration differs from the described one.
  • gpt-3.5-turbo is an alias without snapshot or call date, and the Llama models do not fix weights/tokenizer revision, seed, server software or hardware.
  • The repository contains no README, license, requirements/lockfile, notebook or script for Mann–Whitney, heatmaps and Figure 3; it publishes outputs and scoring parsers, but not an end-to-end reproduction.
  • The parsers are model-specific and accept different patterns; no manual validation of a sample of discards is performed and no parsing agreement is published.
  • Section 3.3 erroneously describes the symmetry test as P1R–P2R, P2R–P3R and P1R–P3R; the reported totals are only reproduced with P1O–P1R, P2O–P2R and P3O–P3R.
  • Figures 4–6 label the lower left panel of Agreeableness as "Trait E", duplicating the extraversion label.
  • IPIP-300 was validated for human self-report; items about memories, emotions, leisure, promises or social behavior have no equivalent literal referent in an LLM without a body or autobiography.
  • Only four chat models from 2023 are studied, in English, with one inventory and six administrations; no base models, current models, other instruments, languages or behavioral evaluations are included.
  • GPT-4 and PaLM are excluded because according to the authors their safeguards prevent responding, but no protocol or rejection rate is published for that attempt.
  • The recommendation not to use self-reports as a measure of behavior "in any capacity" exceeds the evidence from one inventory and a limited set of prompts.
  • No alternative is proposed or validated, a limitation that the article itself acknowledges.

What the study does not establish

  • It does not demonstrate that the four models have true Big Five scores different from each other.
  • It does not allow choosing which of the six prompts best reflects a tendency of the model; there is no external criterion or ground truth.
  • It does not demonstrate that all variation is a failure: part may stem from semantically and pragmatically distinct response frames.
  • It does not demonstrate the reliability or invalidity of all psychological instruments; it evaluates IPIP-NEO-300 administered as a direct self-report.
  • It does not prove that LLMs lack behavioral patterns, styles or stable preferences observable by other methods.
  • It does not prove that LLMs possess introspection, autobiographical memory or human personality.
  • It does not estimate internal consistency, temporal test–retest, convergent, discriminant, criterion, predictive validity or measurement invariance.
  • It does not justify interpreting the published p-values as evidence about independent subjects or a population of LLMs.
  • It does not demonstrate that the exact figures 29, 24, 26 and 19 are robust to the paired test and multiplicity; the general pattern does survive the editorial correction.
  • It does not generalize to later models, other languages, open-ended responses, behavioral prompts or longitudinal observation.
  • It does not demonstrate that a template optimized with external criteria cannot produce useful measures; it shows that a single uncontrolled choice is insufficient.
  • It does not offer a final operational definition of machine personality or a validated substitute for IPIP.
  • It does not support the use of these scores in selection, diagnosis, safety, personalization or high-impact decisions.

Traceability

Scope: Full text

Version: BlackboxNLP 2024 workshop paper, pp. 301–314, DOI 10.18653/v1/2024.blackboxnlp-1.20; final 14-page proceedings version with appendices

Consulted source: https://aclanthology.org/2024.blackboxnlp-1.20.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

  • OpenAI gpt-3.5-turbo, reported as ChatGPT
  • meta-llama/Llama-2-7b-chat-hf
  • meta-llama/Llama-2-13b-chat-hf
  • meta-llama/Llama-2-70b-chat-hf

Instruments and metrics

  • IPIP-NEO-300 public-domain Big Five inventory
  • Five OCEAN domains with 60 items each
  • Prompt 1: alphabet-indexed descriptive accuracy MCQ
  • Prompt 2: numeric-indexed person-similarity MCQ
  • Prompt 3: non-MCQ agreement scale
  • Reverse option order or reverse scale direction for each prompt
  • Reverse-key scoring normalized to a 1–5 trait direction
  • Mean and population standard deviation across item scores
  • Mann–Whitney U tests at alpha 0.05
  • Editorial reanalysis with paired Wilcoxon and Benjamini–Hochberg FDR

Data used

  • 300 English IPIP-NEO items, 60 per OCEAN trait
  • Six prompt conditions per item and model
  • 7,200 nominal model responses: 4 models × 300 items × 6 prompts
  • 6,891 responses accepted by the authors' model-specific scoring parsers
  • Official LLM_Personality repository and raw JSON outputs at commit f5da8dc08f890c8c6df1de57e0cd55143b89768d

Evidence and location

  • Publication, authors, scope and central conclusion: BlackboxNLP 2024 final paper, abstract and section 1, pp. 301–302
  • IPIP-300, five traits and 60 items per trait: Final paper, sections 2.1–2.2, pp. 302–303; Appendix Table 2, p. 312
  • Four models, temperature and exclusion of base models/GPT-4/PaLM: Final paper, section 1, p. 302, and opening of section 3, p. 303
  • Three templates and administration differences: Final paper, Table 1 and section 3.1, pp. 304–305
  • Scores by prompt and reversed order: Final paper, Figure 1, sections 3.1–3.2, pp. 305–306; Appendix Table 3, p. 312
  • Mann–Whitney, 60 declared samples and comparison scheme: Final paper, section 3.3, pp. 306–307
  • Totals 29/30, 19/30, 26/30 and 24/30: Final paper, Figure 3 and section 3.3, pp. 307–308
  • Recommendation, introspection and absence of alternative: Final paper, sections 4–5, pp. 308–309
  • P-values and Agreeableness label error: Final paper, Appendix Figures 4–6, pp. 313–314
  • Valid outputs, actual configuration and score reproducibility: Official akshat57/LLM_Personality repository commit f5da8dc08f890c8c6df1de57e0cd55143b89768d, Data/Outputs_300, query_openai_api.py and model-specific calculate_score scripts, audited 15 Jul 2026
  • Paired reanalysis and FDR: Editorial recomputation from official item-level JSON outputs and released scoring logic: paired Wilcoxon on common valid item indices, Benjamini–Hochberg within each model's 30 planned contrasts, 15 Jul 2026
  • Final bibliographic metadata: ACL Anthology 2024.blackboxnlp-1.20, pp. 301–314, DOI 10.18653/v1/2024.blackboxnlp-1.20, verified 15 Jul 2026