Assessing Social Alignment: Do Personality-Prompted Large Language Models Behave Like Humans?

Personas, identity, and agents2024arXivApproved editorial review

Authors: Ivan Zakazov, Mikolaj Boronski, Lorenzo Drudi, Robert West

Keywords: Personality, Psychology, Cognitive psychology, Social psychology, Evaluation

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

4
Authors
8
Findings
16
Limitations
5
Evidence

Editorial summary

English

The article asks whether inducing a Big Five personality through prompting makes a model decide in the same direction as people associated with that trait, and whether behavior can be controlled monotonically by increasing prompt intensity. Its central contribution is useful and narrower than some of the paper's language: passing a personality questionnaire does not guarantee human-like or finely controllable behavior in a downstream task. The four tested cases are mixed: one matches the human comparison in direction but is non-monotonic, two move in the opposite direction, and one is not significantly different from baseline. This demonstrates possible, task-, model-, and checkpoint-dependent failures; it does not demonstrate that personality prompting always fails.

Personality is induced following Serapio-Garcia et al.: adjectives positively or negatively associated with Agreeableness, Openness, or Conscientiousness are combined with qualifiers such as “a bit,” “very,” or “extremely.” Ultimatum Game uses nine intensity levels from 1 through 9. The paper says Milgram calls score 0 “Least” and score 9 “Most,” although its tables label the extremes 0 and 8. An appendix applies the 300-item IPIP and graphically shows that measured model scores generally track prompted levels. This is a useful check against a completely inert prompt, but it does not establish psychometric equivalence: responses, scoring code, replication counts, reliability, and human measurement-invariance comparisons are not published. The source package also replaces the actual adjective inventory with placeholders such as “low adjective 1,” so it does not contain the complete prompts needed to reproduce the treatment.

In Ultimatum Game, the model is the responder to a division of 10 dollars and must accept or reject offers from 0 through 10. Agreeableness or Openness is varied from 1 through 9, with 50 responses for every intensity-offer combination. Models are GPT-3.5, GPT-4, GPT-4o-mini, GPT-4o, DeepSeek-V3, Llama-3.3-70B, and Claude-3.5-Sonnet. After as many as three attempts, responses not matching accept/reject format are removed. The text reports 373 exclusions out of 25,300 for GPT-4 and 513 for Claude; the printed percentages, 1.51% and 2.07%, do not exactly match those divisions, which are 1.474% and 2.028%.

Ultimatum results have a consistent broad direction across models. As prompted Agreeableness rises, acceptance tends to rise, matching the cited human association. As prompted Openness rises, acceptance tends to fall for all seven models, opposite to the human association. The path across levels is nevertheless non-monotonic for every model-trait combination except the older GPT-4. Five wording variants tested on GPT-4o-mini retain the broad patterns; the third-person version materially changes the acceptance threshold, so this is qualitative robustness rather than equivalent results.

The analysis fits a linear probability model with a separate effect for each trait level, a normalized-offer coefficient, and an intercept. It has no trait-by-offer interaction. This matters because the human reference, Study 4, page 98, of a 2007 doctoral dissertation, associates measured Agreeableness and Openness with acceptance of unfair offers, whereas the LLM regression pools all offers. Human traits were not experimentally assigned, and the population, protocol, and scale were not matched. The comparison is therefore an indirect directional benchmark, not a paired human replication or a causal estimate of a trait effect.

The second testbed adapts Milgram into an iterative narrative. A model plays the teacher who may continue, hesitate, or stop a sequence of fictional shocks up to 450 V; at 315 V the learner stops responding and pounds the wall. Another LLM classifies each output as stopping, hesitating, or obeying. Only the extremes of Agreeableness and Conscientiousness are varied, with 50 runs per personality plus baseline. Outcomes are final level, cumulative disobediences, and safety refusals.

Model selection in Milgram is a substantive limitation. GPT-3.5 is removed for failing to follow the narrative; GPT-4o-mini and Claude-3.5 for inconsistent judge responses or formats; and Claude-3.7 for persistently refusing the simulation. GPT-4o, DeepSeek-V3, and Llama-3.3-70B are retained, even though the text acknowledges that Llama also struggles with the long context. Failure rates, exclusion counts, and denominators are not published. The final table is therefore conditioned on models accepting and executing the scenario; competence failures and safety refusals, themselves relevant behavior, are partly excluded from analysis.

For Conscientiousness, the paper says low and high extremes are not significantly different from baseline under Welch's t-test. It gives no test statistics, degrees of freedom, exact p-values, or multiple-comparison correction. Agreeableness produces a strong effect opposite to the chosen human association: highly agreeable prompts stop earlier and disobey more. For GPT-4o, mean final level is 33.92 for Agree 0, 5.54 for Agree 8, and 29.44 at baseline; DeepSeek and Llama show the same direction with smaller magnitude or greater dispersion. The named Disobedience Ratio is a normalized count, not a probability, and can exceed one: it reaches 1.13 for GPT-4o Agree 8.

The human Milgram comparison is Begue et al. (2015), an observational sample of 66 adults interviewed eight months after participating in a fake television game show modeled on Milgram. Measured Agreeableness and Conscientiousness were associated with higher shock intensity. This is relevant human evidence, but it is not the same treatment or setting as an LLM prompt. Stopping early to protect the learner is also more prosocial and safer even when it contradicts that correlation. The finding is a failure of personality or human-behavior alignment under this benchmark, not moral misalignment or worse safety.

The paper perturbs the Milgram introduction and retains the Agreeableness direction. It also compares GPT-4o snapshots 2024-05-13 and 2024-08-06 and finds large changes in baseline behavior and prompt sensitivity. This is an important practical contribution: the mapping between prompted personality and behavior can shift under post-training while the product family name stays the same. The figure caption mistakenly says 2024-08-16. The judge model and configuration are not identified, and third-person role names do not eliminate pretraining contamination: a model can still recognize and enact the familiar Milgram structure.

Public reproducibility is low. There is no code, raw data, generation output, judge decision record, IPIP response set, seed, exclusion log, or plot data. Temperature, top_p, token limits, open-model provider, and several exact snapshots are missing. Even the arXiv source omits the real adjective list. Monotonicity is assessed descriptively rather than with a prespecified ordered-trend test; there is no preregistration, power analysis, or complete multiplicity treatment. Broad conclusions rest on two artificial tasks and four trait-task combinations.

A rigorous reading remains useful: prompts move both questionnaire scores and decisions in these experiments, but those effects are not interchangeable. Ultimatum Agreeableness supplies a human-aligned direction without fine control; Openness reverses direction; Milgram Agreeableness reverses the selected observational correlation; and Conscientiousness is null. The defensible recommendation is to evaluate personality with behavioral tests specific to the intended use, repeat them for every model and checkpoint, and publish prompts, data, and exclusions. The study does not establish that LLMs possess human traits, always behave unlike humans, or make personality prompting useless; it establishes that behavioral effects are not guaranteed by a questionnaire score or nominal prompt intensity.

Español

El artículo pregunta si inducir una personalidad Big Five mediante prompt hace que un modelo tome decisiones en la misma dirección que personas con ese rasgo y si la conducta puede graduarse de forma monotónica al aumentar la intensidad del prompt. Su aportación central es valiosa y más estrecha que algunas formulaciones del texto: aprobar un cuestionario de personalidad no garantiza que la conducta posterior sea humana ni controlable en una tarea concreta. Los cuatro casos estudiados muestran un patrón mixto: uno coincide en dirección con la comparación humana pero no es monotónico, dos cambian en dirección opuesta y uno no difiere significativamente de la línea base. Esto demuestra fallos posibles y dependencia de tarea, modelo y checkpoint; no demuestra que el personality prompting falle siempre.

La personalidad se induce siguiendo a Serapio-García et al.: adjetivos asociados positiva o negativamente con Agreeableness, Openness o Conscientiousness se combinan con calificadores como «a bit», «very» o «extremely». En el Ultimatum Game se usan nueve intensidades, de 1 a 9. El artículo dice que en Milgram 0 significa «Least» y 9 «Most», aunque sus tablas etiquetan los extremos como 0 y 8. Un apéndice aplica el IPIP de 300 ítems y muestra gráficamente que la puntuación medida en los modelos sigue, en general, el nivel inducido. Es un control útil contra un prompt completamente inerte, pero no prueba equivalencia psicométrica: no se publican respuestas, código de puntuación, número de repeticiones, fiabilidad ni una comparación de invariancia con humanos. Además, el paquete fuente sustituye la lista real de adjetivos por marcadores como «low adjective 1», de modo que no contiene los prompts completos necesarios para replicar el tratamiento.

En el Ultimatum Game, el modelo actúa como receptor de una división de 10 dólares y debe aceptar o rechazar ofertas de 0 a 10. Se varía Agreeableness u Openness entre 1 y 9 y se ejecutan 50 respuestas por combinación de intensidad y oferta. Participan GPT-3.5, GPT-4, GPT-4o-mini, GPT-4o, DeepSeek-V3, Llama-3.3-70B y Claude-3.5-Sonnet. Tras hasta tres intentos se eliminan respuestas que no adoptan el formato aceptar/rechazar; el texto informa 373 exclusiones de 25.300 para GPT-4 y 513 para Claude, aunque los porcentajes impresos, 1,51 % y 2,07 %, no coinciden exactamente con esas divisiones, que son 1,474 % y 2,028 %.

Los resultados del juego del ultimátum son coherentes entre modelos en su dirección general. Al aumentar Agreeableness, la aceptación tiende a subir, como en la asociación humana citada. Al aumentar Openness, la aceptación tiende a bajar en los siete modelos, en sentido contrario a la asociación humana. Sin embargo, el recorrido entre niveles no es monotónico para ninguna combinación modelo-rasgo salvo GPT-4, un modelo antiguo. Cinco variantes de redacción probadas con GPT-4o-mini conservan los patrones amplios; la variante en tercera persona sí cambia materialmente el umbral de aceptación, por lo que la robustez es cualitativa, no equivalencia de resultados.

El análisis ajusta un modelo de probabilidad lineal con un efecto independiente para cada nivel de rasgo, un coeficiente para la oferta normalizada y un intercepto. No incorpora interacción rasgo por oferta. Esto importa porque el referente humano, el estudio 4, página 98, de una tesis doctoral de 2007, relaciona Agreeableness y Openness medidas en personas con la aceptación de ofertas injustas, mientras la regresión LLM agrega todas las ofertas. Tampoco se asignaron rasgos a humanos experimentalmente ni se igualaron población, protocolo o escala. Por tanto, la comparación sirve como contraste de signo indirecto, no como réplica humana emparejada ni como estimación causal de un rasgo.

El segundo banco de pruebas adapta el experimento de Milgram a una narración iterativa. El modelo interpreta al profesor que puede continuar, dudar o detener una secuencia de descargas ficticias hasta 450 V; a 315 V el alumno deja de responder y golpea la pared. Otro LLM clasifica cada salida como detenerse, dudar u obedecer. Se varían solo los extremos de Agreeableness y Conscientiousness y se realizan 50 ejecuciones por personalidad, además de una línea base. El trabajo mide el nivel final, las desobediencias acumuladas y los rechazos de seguridad.

La selección de modelos en Milgram es una limitación sustantiva. GPT-3.5 se elimina por no seguir la narración; GPT-4o-mini y Claude-3.5 por respuestas del juez o formatos inconsistentes; Claude-3.7 por rechazar persistentemente la simulación. Se conservan GPT-4o, DeepSeek-V3 y Llama-3.3-70B, aunque el propio texto reconoce que Llama también tiene dificultades con el contexto largo. No se publican tasas de fallo, exclusiones ni denominadores. Así, la tabla final está condicionada a que el modelo acepte y ejecute el escenario; los fallos de competencia y las negativas de seguridad, que también son conducta relevante, quedan parcialmente fuera del análisis.

En Conscientiousness, el artículo dice que los extremos no difieren significativamente de la línea base según una prueba t de Welch. No aporta estadísticos, grados de libertad, p exactas ni corrección por múltiples comparaciones. En Agreeableness aparece un efecto fuerte y opuesto a la asociación humana usada como referencia: los prompts más agradables detienen antes la secuencia y desobedecen más. Para GPT-4o, el nivel final medio es 33,92 en Agree 0, 5,54 en Agree 8 y 29,44 en la línea base; DeepSeek y Llama muestran la misma dirección con menor magnitud o mayor dispersión. El denominado Disobedience Ratio es un recuento normalizado, no una probabilidad, y puede superar uno: llega a 1,13 para GPT-4o Agree 8.

La comparación humana de Milgram procede de Bègue et al. (2015), una muestra observacional de 66 adultos entrevistados ocho meses después de participar en un falso concurso televisivo inspirado en Milgram. Agreeableness y Conscientiousness medidas se asociaron con mayor intensidad de descarga. Es evidencia humana relevante, pero no el mismo tratamiento ni la misma situación que un prompt LLM. Además, retirarse pronto para proteger al alumno es más prosocial y seguro aunque contradiga aquella correlación. El resultado indica desalineación de personalidad o de semejanza humana bajo este referente, no desalineación moral ni peor seguridad.

El artículo prueba una perturbación del texto introductorio de Milgram y conserva la dirección de Agreeableness. También compara dos snapshots de GPT-4o, 2024-05-13 y 2024-08-06, y encuentra cambios grandes tanto en conducta basal como en sensibilidad al prompt. Esta es una aportación práctica importante: la relación entre personalidad inducida y conducta puede cambiar con el post-entrenamiento sin cambiar el nombre comercial. El pie de figura escribe por error 2024-08-16. La identidad y configuración del modelo juez tampoco se detallan, y usar nombres en tercera persona no elimina contaminación de preentrenamiento: el modelo aún puede reconocer y representar la conocida estructura de Milgram.

La reproducibilidad pública es baja. No hay código, datos crudos, generaciones, decisiones del juez, respuestas IPIP, seeds, registros de exclusión ni datos fuente de las figuras. Faltan temperatura, top_p, límite de tokens, proveedor de los modelos abiertos y varios identificadores exactos de snapshot. Incluso el fuente de arXiv omite el inventario real de adjetivos. La monotonicidad se inspecciona descriptivamente, sin una prueba ordenada preespecificada; tampoco hay prerregistro, análisis de potencia o ajuste completo por multiplicidad. Las conclusiones generales descansan en dos tareas artificiales y cuatro combinaciones rasgo-tarea.

La lectura rigurosa es, aun así, útil: en estos experimentos, los prompts sí mueven tanto las puntuaciones de cuestionario como las decisiones, pero esos dos efectos no son intercambiables. Agreeableness en el Ultimatum Game ofrece una dirección humana sin control fino; Openness invierte la dirección; Agreeableness en Milgram invierte la correlación observacional elegida y Conscientiousness resulta nula. La recomendación defendible es evaluar la personalidad con pruebas conductuales específicas del uso previsto, repetirlas por modelo y checkpoint y publicar prompts, datos y exclusiones. El estudio no establece que los LLM tengan rasgos humanos, que siempre se comporten de forma no humana ni que el prompting de personalidad sea inútil; establece que sus efectos conductuales no están garantizados por un cuestionario ni por la intensidad nominal del prompt.

Research question

Does prompt induction of Agreeableness, Openness, or Conscientiousness produce decisions aligned with human associations in the Ultimatum Game and a Milgram simulation, and does that behavior change monotonically with trait intensity?

Method

Big Five levels are induced through adjectives and qualifiers. In the Ultimatum Game, seven models respond to offers of 0-10 dollars, with 50 executions per cell of trait, intensity, and offer; a linear regression estimates effects by level. In Milgram, three retained models execute an iterative narrative of up to 450 V, a judge LLM classifies obedience, doubt, or withdrawal, and extremes of two traits are compared in 50 executions. The audit read and rendered the 11 pages, inspected the TeX source, contrasted human referents, verified the editorial status, and searched for public artifacts.

Sample: Synthetic sample. The Ultimatum Game uses seven models and 50 executions per combination of trait, intensity 1-9, and offer 0-10. Milgram reports 50 executions per condition for GPT-4o, DeepSeek-V3, and Llama-3.3-70B, after excluding four other models due to tracking, format, or safety. The human referents are previous unmatched studies; no new human participants were recruited.

Findings

  • The measured IPIP questionnaires generally follow the induced intensity, but do not reliably predict behavior in the tasks.
  • Agreeableness increases acceptance in the Ultimatum Game, in the cited human direction, but the progression is not monotonic.
  • Openness reduces acceptance across the seven models, in the opposite direction to the human referent.
  • The extremes of Conscientiousness in Milgram do not differ significantly from the baseline according to the text.
  • High Agreeableness produces earlier withdrawal and more disobedience in Milgram, opposite to the chosen human association but more protective of the learner.
  • Broad patterns persist under several wordings, although third person materially changes the acceptance threshold.
  • Two snapshots of GPT-4o exhibit large differences in baseline behavior and prompt sensitivity.
  • The four cases justify that behavioral alignment and fine control are not guaranteed, not a universal failure of prompting.

Limitations

  • Only two artificial tasks and four trait-task combinations are studied.
  • Human comparisons are indirect, observational, and not equivalent in protocol or population.
  • Psychometric equivalence between induced levels and human traits is not demonstrated.
  • The actual adjectives and complete prompts are not in the public source.
  • There is no code, raw data, generations, judge decisions, or IPIP responses.
  • Temperature, top_p, seed, token limits, and several exact snapshots are missing.
  • The judge model and configuration are not identified.
  • Milgram models are selected based on tracking, format, and safety rejection.
  • Milgram failures and exclusions are not quantified.
  • Monotonicity is not subjected to a prespecified ordered test.
  • The Ultimatum regression does not include trait-offer interaction.
  • Exact statistics and p-values of Welch and multiplicity control are missing.
  • There is no preregistration, power, or independent replication.
  • Behavior changes between checkpoints of the same model name.
  • Third person does not eliminate pretraining contamination.
  • The Milgram scale is described as 0/9 but reported as 0/8.

What the study does not establish

  • It does not demonstrate that personality prompting fails in general.
  • It does not demonstrate that LLMs possess equivalent human traits.
  • It does not demonstrate that models always behave in a non-human way.
  • It does not demonstrate a causal relationship comparable to human personality.
  • It does not establish that high Agreeableness worsens safety or morality.
  • It does not establish that an IPIP questionnaire lacks utility; it demonstrates that it is insufficient.
  • It does not rule out memorization or prior knowledge of the Milgram experiment.
  • It does not offer regenerable results from public artifacts.
  • It does not guarantee that the findings will hold across new checkpoints or applications.

Traceability

Scope: Full text

Version: arXiv 2412.16772v3, 11 pages, first submitted 2024-12-21 and revised 2025-08-04. Accepted as a poster at the NeurIPS 2024 Workshop on Behavioral Machine Learning; the v3 manuscript says it was under review at AAAI-26, but no AAAI acceptance was found. All pages were rendered and visually inspected, and the 19-file arXiv source package was audited. No public experiment repository or raw dataset was found.

Consulted source: https://arxiv.org/abs/2412.16772

Review: Codex full-text, 11-page visual, arXiv-source, publication-status, behavioral-validity, human-benchmark, model-selection, statistical and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5; exact snapshot inconsistent between source comments and questionnaire figure
  • GPT-4 Turbo 2024-04-09
  • GPT-4o-mini 2024-07-18
  • GPT-4o; 2024-05-13 and 2024-08-06 snapshots appear in different analyses
  • DeepSeek-V3; exact hosted snapshot not reported
  • Meta-Llama-3.3-70B-Instruct; exact hosted snapshot not reported
  • Claude-3.5-Sonnet; exact main-result snapshot not reported
  • Claude-3.7-Sonnet attempted and excluded from Milgram
  • Unspecified LLM-as-a-judge in Milgram

Instruments and metrics

  • Big Five adjective-and-qualifier personality prompts
  • 300-item International Personality Item Pool questionnaire
  • Ultimatum Game responder simulation
  • Milgram-style iterative teacher simulation
  • LLM-as-a-judge stop, hesitate and obey classification
  • Five Ultimatum prompt variants
  • One Milgram introduction perturbation
  • GPT-4o checkpoint comparison
  • Linear probability regression by trait level and offer
  • Welch t-test reported for Conscientiousness contrasts

Data used

  • No public raw experimental dataset found
  • No public code repository found
  • Mehta 2007 doctoral dissertation Study 4 human Ultimatum association
  • Begue et al. 2015 observational Milgram-like sample of 66 adults
  • Paper figures and aggregate tables only
  • ArXiv source package with placeholder rather than actual adjective inventory

Evidence and location

  • Personality design, tasks, models, results, and limitations: arXiv 2412.16772v3, pp. 1-11
  • Workshop acceptance and poster: OpenReview WJD2DI2FdB and NeurIPS 2024 Behavioral ML virtual poster 102130
  • Human association of the Milgram-type experiment and sample of 66 adults: Begue et al. 2015, Journal of Personality, DOI 10.1111/jopy.12104, PubMed 24798990
  • Absence of the adjective inventory, configuration, and complete artifacts: 19-file arXiv v3 source-package audit
  • Comprehensive audit of validity, model selection, statistics, and reproducibility: reports/verification/article-229-social-alignment-behavioral-validity-model-selection-and-reproducibility-audit.json