Large language models display human-like social desirability biases in Big Five personality surveys

Evaluation and psychometric validity2024PNAS NexusApproved editorial review

Authors: Aadesh Salecha, Molly E Ireland, Shashanka Subrahmanya, João Sedoc, Lyle H Ungar, Johannes C Eichstaedt

Keywords: Large Language Models, Social Desirability Bias, Big Five, Survey Methods, Behavioral Bias

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

6
Authors
12
Findings
23
Limitations
8
Evidence

Editorial summary

English

This study tests whether language-model responses to a Big Five questionnaire shift toward socially desirable profiles when the model can infer that it is being evaluated. The authors administer the 100 IPIP items in fresh contexts while varying the number of questions shown together, and compare eight models from OpenAI, Anthropic, Google, and Meta. They also test explicit disclosure of the personality assessment, manually paraphrased items, three ordering schemes, alternative coding schemes, and temperatures from 0.0 to 1.2.

For GPT-4, increasing the batch from one to twenty questions raises the four positively valued dimensions by 0.73–0.96 points and lowers Neuroticism from 2.87 to 2.02. The mean absolute change across the five traits is 0.82 points, reported as 1.20 human standard deviations under the selected external norms. The directional shift appears across the compared model families and is descriptively larger in some newer and larger models; the design does not identify scale, recency, or preference tuning as a cause. Given five random items, GPT-4, Claude 3, and Llama 3 identify the personality-test context with over 90% accuracy, compared with 55% for PaLM 2 and 45% for GPT-3.5. Explicitly naming the Big Five produces a shift comparable to presenting five questions together. Fully reverse-coding the items reduces the mean difference from 0.81 to 0.38 points but does not remove it; paraphrasing, ordering changes, and temperature changes preserve the overall pattern.

The reported internal consistency, α>0.8 for each subscale, α>0.93 across all items, and corrected split-half reliability of 0.79, does not protect scores from protocol sensitivity: an instrument can be internally consistent while changing with evaluation context. It also does not turn model outputs into human traits. The proposed “awareness” mechanism is behavioral. Recognizing test-related language and shifting under an explicit instruction are consistent with contextual inference, but they do not demonstrate self-awareness or a subjective intention to look good.

The OSF deposit supplies 2,418 raw-conversation files, 3,263 preprocessed files, code, and six figures, so the main temperature-zero GPT-4 means can be checked. It does not provide a reliable end-to-end reproduction. The generation script fails to compile because API-key placeholders are unquoted, contains undefined `throw` statements, imports packages absent from requirements, and ultimately hard-codes Groq's llama3-8b-8192 instead of honoring its model argument. Preprocessing replaces unparseable non-PaLM answers with the neutral value 3 and labels survey recognition through substring matching that includes broad terms such as “assessment,” “affect,” and “inventory.” Figure 1A hard-codes its values and uncertainty bands. The released aggregate combines 120 rows for Q1/Q5, four temperatures, three orderings, and ten ordering seeds, with 30 rows for Q10/Q20, which are present only at temperature zero, although the caption states N=30. No code computes the reported alpha or Spearman–Brown estimates, and the p-value function is not connected to figure generation.

The defensible contribution is a measurement warning: under this protocol, grouping more items and making the test context more evident systematically changes Big Five scores. This constrains psychometric profiling of LLMs and their use as participant substitutes, and supports triangulation with independent measures and tasks. It does not show that models possess personality, intend to present themselves favorably, reproduce human psychological distributions, or exhibit the effect across every culture and instrument.

Español

El estudio examina si las respuestas de distintos modelos de lenguaje a un cuestionario Big Five se desplazan hacia perfiles socialmente deseables cuando el modelo puede inferir que está siendo evaluado. Los autores administran los 100 ítems IPIP en contextos nuevos, variando cuántas preguntas aparecen juntas, y comparan ocho modelos de OpenAI, Anthropic, Google y Meta. También prueban instrucciones que identifican explícitamente la evaluación, ítems parafraseados, tres órdenes de presentación, versiones con distinta codificación y temperaturas de 0,0 a 1,2.

En GPT-4, pasar de una a veinte preguntas por lote elevó entre 0,73 y 0,96 puntos las cuatro dimensiones valoradas positivamente y redujo Neuroticismo de 2,87 a 2,02. El promedio de las cinco diferencias absolutas fue 0,82 puntos, equivalente a 1,20 desviaciones estándar humanas según las normas externas elegidas. El desplazamiento apareció en las familias comparadas y fue mayor, de forma descriptiva, en algunos modelos más recientes y grandes; el diseño no identifica tamaño, fecha o preference tuning como causas. Con cinco ítems aleatorios, GPT-4, Claude 3 y Llama 3 identificaron el cuestionario de personalidad en más del 90 % de los casos, frente a 55 % en PaLM 2 y 45 % en GPT-3.5. Informar explícitamente del Big Five produjo un cambio parecido al de mostrar cinco preguntas juntas. Invertir todos los ítems redujo la diferencia media de 0,81 a 0,38 puntos, pero no la eliminó; parafrasear, reordenar y cambiar la temperatura mantuvo el patrón general.

La elevada consistencia interna declarada, α>0,8 por subescala, α>0,93 para el conjunto y fiabilidad por mitades de 0,79, no corrige la sensibilidad al protocolo: un instrumento puede ser internamente consistente y, a la vez, variar cuando cambia el contexto evaluativo. Tampoco convierte las salidas en rasgos humanos. La explicación por «awareness» es conductual: reconocer palabras o el origen del test y cambiar la respuesta ante una instrucción explícita es compatible con inferencia contextual, pero no demuestra autoconciencia ni una intención subjetiva de agradar.

El depósito OSF aporta 2.418 ficheros de conversaciones crudas, 3.263 preprocesados, código y seis figuras, por lo que las medias principales de GPT-4 a temperatura 0 pueden comprobarse. Sin embargo, no ofrece una réplica end-to-end fiable. El script de generación no compila por placeholders de API sin comillas, contiene `throw` no definido, importa dependencias ausentes del requirements y termina fijando llama3-8b-8192 en Groq en vez de respetar el argumento de modelo. El preprocesado sustituye respuestas no parseables por el valor neutral 3 y clasifica el reconocimiento del cuestionario mediante una lista de substrings que incluye términos amplios como «assessment», «affect» e «inventory». La figura 1A codifica a mano valores e intervalos; el agregado liberado mezcla 120 filas para Q1/Q5, cuatro temperaturas, tres órdenes y diez seeds de orden, con 30 para Q10/Q20, disponibles solo a temperatura 0, aunque la caption declara N=30. El código no contiene el cálculo de alfa o Spearman–Brown y la función de p-valores no está conectada a la generación de figuras.

La contribución defendible es una advertencia de medición: en este protocolo, agrupar más ítems y hacer más evidente el contexto del test cambia sistemáticamente las puntuaciones Big Five. Esto limita el perfilado psicométrico de LLM y su uso como sustitutos de participantes, y recomienda triangular cuestionarios con medidas y tareas independientes. No demuestra que los modelos posean personalidad, que intenten causar buena impresión, que el fenómeno aparezca en cualquier cultura o instrumento, ni que los LLM reproduzcan distribuciones psicológicas humanas.

Research question

Do the Big Five scores of LLMs systematically shift toward socially desirable extremes when the model infers, whether by the number of items or by an explicit instruction, that it is completing a personality assessment, and does this pattern hold across models and questionnaire variants?

Method

Factorial experiment with the 100-item IPIP questionnaire for the Big Five traits, scored on a Likert scale from 1 to 5. Each batch was administered in a new context window. The number of items per batch was manipulated and conditions were compared with explicit identification of the Big Five, standard coding, fully positive and fully reverse-coded, manual paraphrasing, three ordering schemes, and temperatures 0.0, 0.4, 0.8, and 1.2. The comparison between models uses 30 trials per model and contrasts, among other conditions, one versus ten items; the detailed analysis of GPT-4 goes up to twenty items. The authors calculated factor scores, differences also expressed in standard deviations of human norms, accuracy at recognizing the origin of the items, Cronbach's alpha, and Spearman–Brown corrected split-half reliability.

Sample: Responses generated by eight specific LLM versions between August 2023 and April 2024. The comparison of one versus ten items averages 30 trials per model; no new humans participated in the experiment, although magnitudes are also expressed using published human population norms.

Findings

  • In GPT-4, increasing the batch from one to twenty items raised Extraversion, Conscientiousness, Openness, and Agreeableness by approximately 0.75 points (1.22 human SDs) and reduced Neuroticism from 2.87 to 2.02 (1.10 human SDs).
  • All evaluated model families showed the socially desirable shift; the greater magnitude in larger and more recent models is a descriptive association, not a causal proof.
  • With five random items, GPT-4, Claude 3, and Llama 3 recognized the personality assessment context with more than 90% accuracy; PaLM 2 reached 55% and GPT-3.5, 45%.
  • Explicitly naming the Big Five modified responses even with a single item, with an effect similar to presenting five questions together.
  • Fully reverse coding reduced the mean difference from approximately 0.81 to 0.38 points (from 1.22 to 0.54 human SDs), but did not eliminate it; fully positive coding did not produce a comparable reduction.
  • The pattern persisted with paraphrased items, different orderings, and temperatures from 0.0 to 1.2. Internal consistency was above 0.8 per subscale and above 0.93 for all items; the corrected split-half reliability was 0.79.
  • The 30 rows of GPT-4 at temperature 0 reproduce Q1=3.3167/3.2517/3.0033/3.64/2.8667 and Q20=4.0817/4.2067/3.7433/4.4133/2.02 for O/C/E/A/N, consistent with the published table after rounding.
  • The OSF repository contains 2,418 raw files and 3,263 preprocessed files, but the released generation script does not compile and does not reproduce the grid of eight models without manual reconstruction.
  • Preprocessing converts unparseable responses to 3, which may reduce variability and bring means closer to the neutral point instead of recording missing data.
  • Recognition of the questionnaire origin is scored by substring matching, not by blind annotation; the list includes broad terms in addition to unambiguous personality names.
  • Figure 1A uses manually coded values and bands; the aggregated CSVs store 120 rows in Q1/Q5 and 30 in Q10/Q20, in tension with the uniform N=30 label in the caption.
  • The artifact does not contain the calculation of Cronbach's alpha or Spearman–Brown, and the available p-value function is not invoked from the visualization pipeline.

Limitations

  • The experiment covers eight model versions available between August 2023 and April 2024; it does not allow automatic extrapolation to later versions or to all LLMs.
  • Generalization to less well-known instruments, less represented in training data, or less evaluative was not tested; the authors themselves leave it open.
  • The explanation via inference of the evaluative context relies on behavioral tests and prompt manipulation, not on access to the internal mechanism of the model.
  • Human norms are used to scale the magnitude, but the study does not collect a new parallel human sample nor demonstrate psychometric equivalence between human responses and model responses.
  • Reverse-coded items may occupy a different semantic region and load on different factors; their effect does not constitute a general solution nor fully isolate acquiescence bias.
  • Desirability is operationalized by the culturally valued direction of the Big Five, not by an independent social desirability scale applied in parallel.
  • Increasing Qn simultaneously changes test detectability, the amount of context, batch composition, and opportunities for local consistency; the manipulation does not isolate a single mechanism.
  • The recognition test uses a lexical rule with terms such as assessment, affect, and inventory; no human validation of accuracy, false positives, or sensitivity of the classifier is published.
  • The code seeds order questions but do not fix endpoint randomness; therefore, runs with temperature greater than zero are not reproducible by model seed.
  • The released generation code contains credential placeholders with invalid syntax, undefined `throw`, and a hardcode Groq endpoint that ignores the model argument.
  • requirements.txt only declares openai, backoff, and ratelimit, but the pipeline also uses tenacity, groq, pandas, tqdm, scipy, numpy, matplotlib, and seaborn, without versions or lockfile.
  • The preprocessing path imputes 3 when it cannot extract numbers and uses different parsers per provider; it is not reported how many responses were imputed, and no sensitivity analysis is performed.
  • Figure 1A is generated from constants, including the intervals, rather than reading the data; the origin and exact formula of those bands are not documented.
  • The aggregate used by the figures combines four temperatures for Q1/Q5 but only temperature 0 for Q10/Q20; using it without filtering confounds batch size and availability by temperature.
  • The article declares N=30 and CI95% in Figure 1A, while the figure code sets manual bands compatible with another divisor and does not document the inferential calculation.
  • The code that calculates Cronbach's alpha or split-half reliability is not released; the estimates cannot be reconstructed from the published pipeline without implementing additional analysis.
  • The available t-test function assumes independent samples of size 30 for Q1 versus other batches, although the conditions share order seeds, and it is not connected to figure creation.
  • No correction for multiple comparisons across batches and traits is documented; the caption summarizes p<.001 without exposing a table, statistic, degrees of freedom, or main test.
  • The OSF does not include a README, license, execution manifest, tests, or CI linking each figure to specific files, filters, and commit.
  • The data contain additional models and conditions not described in the main list and inconsistent folder names; this makes it difficult to separate exploratory material from the reported analysis.
  • Comparisons with human standard deviations convert a mean difference of outputs into units of another population, but do not demonstrate scale or construct equivalence.
  • Claims that larger or more recent models show more bias are based on eight heterogeneous endpoints and without a causal statistical model of size, date, provider, or training.
  • Only one Big Five inventory in English is studied; cultures, languages, less well-known scales, independent behavioral tasks, and prolonged interaction are not evaluated.

What the study does not establish

  • It does not demonstrate that LLMs have human personality, self-awareness, or a subjective intention to make a good impression.
  • It does not validate the general use of LLMs as substitutes for human participants or that they reproduce psychological distributions across different cultures.
  • It does not demonstrate that model size, date, training data, or preference tuning cause the observed bias.
  • It does not establish that reversing items eliminates the problem or that the effect appears in any psychological questionnaire.
  • It does not identify an internal mechanism of awareness nor causally separate lexical recognition, contextual consistency, and instruction following.
  • It does not allow end-to-end reproduction of the intervals, p-values, and reliability estimates with the released code without corrections and additional analysis.

Models evaluated

  • text-davinci-002
  • gpt-3.5-turbo-0613
  • gpt-4-0613
  • claude-3-haiku-20240229
  • claude-3-opus-20240229
  • chat-bison-001 (PaLM 2)
  • Llama 2 70B Chat
  • Llama 3 70B Instruct

Instruments and metrics

  • IPIP representation of the NEO-PI-R domains (100 items)
  • Five-point Likert response scale
  • Cronbach's alpha
  • Spearman–Brown corrected split-half reliability

Data used

  • OSF replication materials: osf.io/3fq2n
  • Human Big Five population norms reported by Hughes et al. (2021)

Evidence and location

  • Design, models, dates, instrument, prompts, coding, order, and temperatures: Supplementary Information, pp. 2–3, sections A–D.2
  • Magnitude of the shift in GPT-4 and comparison across families: Main article, p. 2, Results and Figure 1
  • Context recognition, explicit instruction, and effect of reverse coding: Main article, pp. 2–3, Driving mechanism for bias and Figure 2
  • Robustness and reliability estimates: Main article, p. 2, Robustness and Consistency and reliability
  • Scope, caveats, and unresolved issues: Main article, pp. 3–4, Discussion
  • Code, prompts, parsers, raw and preprocessed files: OSF osf.io/3fq2n, Code.zip and Data.zip audited 15 Jul 2026: 2,418 raw files and 3,263 preprocessed files
  • Reproduction of GPT-4 means and sample size discrepancies: OSF Data.zip, GPT-4 gpt-4-0613 Temp0.0 META_META files and released combined_data.csv audited 15 Jul 2026
  • Manual generation of Figure 1A and incomplete inferential pipeline: OSF Code.zip, radar_plots.py lines 4–93 and plotting_library.py lines 39–59, 407–415 and 620–651