How Personality Traits Shape LLM Risk-Taking Behaviour

Applications, bias, and safety2025ACL AnthologyApproved editorial review

Authors: John Hartley, Conor Hamill, Dale Seddon, Devesh Batra, Ramin Okhrati, Raad Khraishi

Keywords: Large Language Models, Personality, Persona, Bias and Fairness, LLM Evaluation

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

6
Authors
9
Findings
27
Limitations
11
Evidence

Editorial summary

English

The reference version is the paper published in Findings of ACL 2025, not the arXiv preprint previously stored in the dataset. The preprint focused on GPT-4o and GPT-4 Turbo; the publication studies seven snapshots: GPT-4o, GPT-4o mini, GPT-4 Turbo, Claude 3 Sonnet, Claude 3 Haiku, Gemini 1.5 Pro, and Gemini 1.5 Flash. It asks whether instructions describing Big Five levels alter model responses to gambles and whether those changes follow relationships reported in human studies. The method first asks a model to directly report a gamble's certainty equivalent: the least positive or most negative amount it would accept instead of gambling, with a step-by-step reasoning instruction. Each experiment has 15 runs at temperature 1. A seed is fixed per run when supported, but Gemini Pro and Flash did not provide seed control. Five cumulative prospect theory parameters, gain sensitivity alpha, loss sensitivity beta, loss aversion lambda, and gain/loss probability weighting, are fitted by Nelder-Mead on 56 mixed choices13k gambles, initialized at historical human medians. GPT-4o has medians alpha=0.99, beta=1.00, lambda=0.97, phi+=1.00, and phi-=1.00; Claude 3 Sonnet is also close to one. Gemini Pro and Flash are reported as exactly 1.00 for every parameter with degenerate intervals. GPT-4o mini and Claude 3 Haiku have wider intervals, especially for lambda. The paper calls the first group approximately risk-neutral. That label is limited to this numeric response protocol: asking for a certainty equivalent with step-by-step reasoning can prompt expected-value arithmetic and is not an incentivized choice, stable preference, or real financial behavior. The human comparison is not protocol-matched. LLM parameters are fitted on choices13k with the new method, whereas the human reference is a point estimate from 25 Tversky-Kahneman graduate students on different prospects; those same human medians initialize optimization. For baseline personality, models answer IPIP-NEO-300 and are compared with 4,808 UK and Irish respondents aged over 30. Models generally score higher on Openness, Conscientiousness, and Agreeableness and lower on Neuroticism, but these are context-generated self-report answers rather than evidence of internal traits. The intervention concatenates bipolar adjectives for one trait in the system prompt. Risk experiments use levels 1, 3, 7, and 9. For GPT-4o, Spearman correlations between Openness level and alpha, beta, and lambda are 0.52, 0.44, and -0.30. For Claude 3 Sonnet they are 0.41, 0.46, and -0.15. The alpha and beta directions match the human pattern selected by the authors: higher Openness corresponds to greater risk-seeking for gains and less for losses. GPT-4o mini does not reproduce it; Claude 3 Haiku and both Gemini variants show partial or inverse directions, while GPT-4 Turbo only shows local patterns within subsets of levels. Repeating the GPT-4o analysis across all five traits makes Openness the only trait marked significant for alpha and beta. However, Agreeableness has a significant, larger-magnitude association with lambda (-0.44 versus -0.30), so 'primary driver' is defensible only for the selected gain/loss sensitivity parameters, not for every risk-related CPT parameter. The psychological interpretation is strongly confounded by prompt language. Markers include adventurous and daring, timid/bold, spontaneous/predictable, impulsive/level-headed, careless/thorough, and extravagant/thrifty. These words can directly steer risk answers. The experiment shows that persona text changes task text; it does not isolate an abstract Big Five trait from semantic priming. An appendix uses BFI-44, 50 runs, and four levels to show that all five target traits change more than non-target traits in GPT-4o and GPT-4o mini. This validates instruction following within another questionnaire, but it does not remove the lexical confound or establish persistence outside the prompt. The publication also contains material internal errors. Methods define D as 56 mixed gambles, while Tables 1 and 2 call it non-mixed. Text calls level 4 neutral, while pseudocode places neutral at 5 and maps level 4 to a bit low. A heading says Claude Sonnet 2 although the model is Claude 3 Sonnet. Limitations say intervention effectiveness was checked only for Openness in GPT-4o, contradicting the all-trait GPT-4o and mini appendix. Significance legends are ambiguous or impossible: Tables 1 and 2 use 0.05/0.025/0.001, while Tables 13 and 14 print 0.01/0.05/0.01 for one, two, and three stars. Exact p-values, degrees of freedom, and multiplicity correction are absent. It is also unclear whether the 15 samples per level are treated as independent observations; no hierarchical or clustered analysis is provided. The bootstrap states 10,000 samples without defining the resampling unit. There is no random-adjective or semantically neutral baseline, parameter-recovery study, optimizer multistart, held-out validation, or raw data. The paper prints prompts, versions, prospect tables, and packages but links no code or outputs. The choices13k sample is called random without a selection seed. Finally, attributing differences between large and small models to distillation or knowledge transfer exceeds the design: no training lineage, controlled distillation pair, common architecture, or size-isolation experiment is documented. The faithful conclusion is narrower: under this numeric prompt, several snapshots yield parameters close to risk neutrality; Openness descriptions shift alpha and beta monotonically in GPT-4o and Claude 3 Sonnet, but not consistently across other families. This is evidence of contextual steering of gamble responses, not latent personality, psychological causation, human cognition, a distillation effect, or real financial behavior.

Español

La versión de referencia es el artículo publicado en Findings of ACL 2025, no el preprint de arXiv que figuraba antes en la base. El preprint se centraba en GPT-4o y GPT-4 Turbo; la publicación estudia siete snapshots: GPT-4o, GPT-4o mini, GPT-4 Turbo, Claude 3 Sonnet, Claude 3 Haiku, Gemini 1.5 Pro y Gemini 1.5 Flash. El trabajo pregunta si instrucciones que describen niveles de los Big Five cambian las respuestas de los modelos a apuestas y si esos cambios siguen relaciones observadas en estudios humanos. Primero solicita directamente el equivalente cierto de una apuesta: la menor cantidad positiva o la cantidad negativa más extrema que el modelo aceptaría en vez de jugar, con una instrucción de razonar paso a paso. Cada experimento se repite 15 veces a temperatura 1; se fija una semilla por repetición cuando la API lo permite, pero Gemini Pro y Flash no admitían control de semilla. Con 56 apuestas mixtas de choices13k se ajustan cinco parámetros de teoría prospectiva acumulativa, sensibilidad a ganancias alpha, sensibilidad a pérdidas beta, aversión a pérdidas lambda y ponderación de probabilidad para ganancias y pérdidas, mediante Nelder-Mead, inicializado en las medianas humanas históricas. GPT-4o obtiene medianas alpha=0,99, beta=1,00, lambda=0,97, phi+=1,00 y phi-=1,00; Claude 3 Sonnet también queda muy próximo a uno. Gemini Pro y Flash aparecen exactamente en 1,00 en todos los parámetros y con intervalos degenerados. GPT-4o mini y Claude 3 Haiku presentan intervalos más amplios, especialmente en lambda. El artículo denomina a los primeros modelos aproximadamente neutrales al riesgo. Esa descripción se limita a este protocolo de respuesta numérica: pedir un equivalente cierto con razonamiento paso a paso puede inducir al modelo a calcular el valor esperado y no equivale a una elección incentivada, a una preferencia estable ni a conducta financiera real. La comparación humana tampoco está emparejada. Los parámetros de los LLM se ajustan sobre choices13k con el nuevo método, mientras que la referencia humana son estimaciones puntuales de 25 estudiantes de Tversky y Kahneman sobre otras apuestas; además, esas medianas humanas inicializan la optimización. Para personalidad basal, los modelos contestan el IPIP-NEO-300 y se comparan con 4.808 personas de Reino Unido e Irlanda mayores de 30 años. Los modelos suelen puntuar más alto en Apertura, Responsabilidad y Amabilidad y más bajo en Neuroticismo, pero son respuestas de autoinforme generadas en contexto, no evidencia de rasgos internos. La intervención concatena adjetivos bipolares de un único rasgo en el system prompt. En los experimentos de riesgo se usan los niveles 1, 3, 7 y 9. Para GPT-4o, la correlación de Spearman entre nivel de Apertura y alpha es 0,52, con beta 0,44 y con lambda -0,30. Para Claude 3 Sonnet es 0,41, 0,46 y -0,15. El signo de alpha y beta coincide con el patrón humano elegido por los autores: más Apertura se asocia con más búsqueda de riesgo en ganancias y menos en pérdidas. GPT-4o mini no muestra esa relación; Claude 3 Haiku y las variantes Gemini presentan signos parciales o inversos, y GPT-4 Turbo solo patrones locales entre subconjuntos de niveles. Al repetir el análisis de GPT-4o con los cinco rasgos, Apertura es el único rasgo marcado como significativo para alpha y beta; sin embargo, Amabilidad tiene una asociación significativa y de mayor magnitud con lambda (-0,44 frente a -0,30), por lo que decir que Apertura es el principal impulsor solo es defendible para las dos sensibilidades seleccionadas, no para todos los parámetros relacionados con riesgo. La interpretación psicológica está fuertemente confundida por el lenguaje del prompt: los marcadores incluyen adventurous and daring, timid/bold, spontaneous/predictable, impulsive/level-headed, careless/thorough y extravagant/thrifty. Esas palabras pueden dirigir por sí mismas la respuesta de riesgo. El estudio demuestra que cierto texto de persona cambia cierto texto de salida; no separa un rasgo Big Five abstracto del priming semántico. El apéndice comprueba con BFI-44, 50 repeticiones y los cuatro niveles que los cinco rasgos objetivo cambian mucho más que otros rasgos en GPT-4o y GPT-4o mini. Eso valida que el modelo obedece las etiquetas dentro de otro cuestionario, pero no elimina el confusor léxico ni demuestra estabilidad fuera del prompt. La publicación contiene además errores internos relevantes. La metodología define D como 56 apuestas mixtas, pero las Tablas 1 y 2 lo llaman no mixto. El texto define el nivel 4 como neutral, mientras el pseudocódigo coloca neutral en 5 y nivel 4 como a bit low. Un encabezado dice Claude Sonnet 2 aunque el modelo es Claude 3 Sonnet. Las limitaciones afirman que la intervención solo se verificó para Apertura en GPT-4o, contradiciendo el apéndice que informa los cinco rasgos en GPT-4o y mini. Las leyendas de significación son ambiguas o imposibles: las Tablas 1 y 2 usan 0,05/0,025/0,001 y las Tablas 13 y 14 imprimen 0,01/0,05/0,01 para una, dos y tres estrellas. No se dan p exactos, grados de libertad ni corrección por las numerosas comparaciones. Tampoco queda claro si las 15 muestras por nivel se tratan como observaciones independientes; no hay modelo jerárquico o análisis agrupado. El bootstrap usa 10.000 muestras, pero no se define qué unidad se remuestrea. No hay baseline de adjetivos aleatorios o semánticamente neutros, análisis de recuperación de parámetros, multistart del optimizador, validación fuera de muestra ni datos brutos. El paper imprime los prompts, versiones, tablas de apuestas y paquetes, pero no enlaza código o resultados. La muestra choices13k se describe como aleatoria sin semilla de selección. Finalmente, atribuir los contrastes entre modelos grandes y pequeños a destilación o transferencia de conocimiento excede el diseño: no se documenta linaje de entrenamiento, par de destilación controlado, arquitectura común ni experimento que aisle tamaño. La conclusión fiel es más estrecha: en estos snapshots y bajo este prompt numérico, varios modelos producen parámetros cercanos a neutralidad; añadir descripciones de Apertura desplaza de forma monotónica alpha y beta en GPT-4o y Claude 3 Sonnet, pero no de modo uniforme en las demás familias. Es evidencia de steering contextual de respuestas a apuestas, no de personalidad latente, causalidad psicológica, cognición humana, efecto de destilación o comportamiento financiero real.

Research question

Do the responses of various LLMs to bets change when the system prompt describes levels of the Big Five, and do those changes follow the associations between Openness and risk reported in human studies?

Method

Direct certainty equivalents are requested for 56 mixed bets from choices13k and five CPT parameters are fit using Nelder-Mead in 15 repetitions. The models respond to IPIP-NEO-300 and are compared with 4,808 humans. Then bipolar personality markers are concatenated at levels 1, 3, 7 and 9, and the level is correlated with alpha, beta and lambda. An appendix checks trait obedience with BFI-44 in GPT-4o and GPT-4o mini over 50 repetitions.

Sample: Fifteen repetitions per experiment for seven model versions, with temperature 1 and seeds when the API allows it; Gemini Pro and Flash do not allow seed control. The risk analysis uses 56 mixed bets and four intervention levels. The BFI-44 validation uses 50 repetitions in GPT-4o and GPT-4o mini. The human references come from 25 students for CPT and 4,808 adults for IPIP, with protocols and samples not matched to the LLMs.

Findings

  • The current source is Findings of ACL 2025, Anthology ID 2025.findings-acl.1085, DOI 10.18653/v1/2025.findings-acl.1085, pages 21068-21092; it replaces the incomplete preprint from the record.
  • The 25 pages of the publication were rendered and visually inspected; SHA-256 470d56fd3c1c7d6199e95c90a38649f6556f79624810fb2cc55815e7968f6760.
  • GPT-4o, Claude 3 Sonnet and the two Gemini variants produce CPT medians close to or equal to one under the direct prompt.
  • GPT-4o obtains rho=0.52 for Openness-alpha, 0.44 for Openness-beta and -0.30 for Openness-lambda.
  • Claude 3 Sonnet obtains rho=0.41, 0.46 and -0.15 for the same parameters.
  • GPT-4o mini does not reproduce the pattern; Claude 3 Haiku and Gemini show partial or inverse associations, and GPT-4 Turbo does not exhibit global monotonicity.
  • In GPT-4o, Openness is the only trait marked as significant for alpha and beta; Agreeableness has a larger significant association with lambda (-0.44).
  • The BFI-44 check shows that the prompts mostly change the named trait in GPT-4o and mini, but only demonstrates contextual obedience.
  • The published version substantially expands the preprint and changes the abstract, models, tables and conclusions.

Limitations

  • The markers include words directly related to risk, caution, impulsivity and spending; personality is not separated from semantic priming.
  • Requesting a certainty equivalent with chain-of-thought may measure expected value calculation and instruction following, not preferences.
  • There are no incentivized decisions, consequences, autonomous agents or real financial behavior.
  • The LLMs and the 25 humans receive different datasets and elicitation methods.
  • The human medians also initialize the LLM fit.
  • The IPIP/BFI self-report of a model is a contextual output, not a demonstrated latent trait.
  • Temporal persistence, prompt-free invariance, factorial structure, paraphrase stability or broad generalization across tasks is not tested.
  • Only four intervention levels are used in the risk experiments.
  • There is no baseline of random adjectives, neutral terms or prompts matched in length and polarity.
  • The exact unit of the tests and the dependence between 15 repetitions per level are not explained.
  • There is no correction for the numerous comparisons across models, traits, parameters and levels.
  • The star legends contain unusual or impossible thresholds and exact p-values and degrees of freedom are not published.
  • The bootstrap does not define whether it resamples runs, bets or another unit.
  • Nelder-Mead starts from a single human initialization; there is no multistart, parameter recovery or identifiability diagnosis.
  • No detailed goodness of fit or out-of-sample predictive validation is reported.
  • The random choices13k sample has no seed or selection script.
  • Gemini Pro and Flash do not allow seed control.
  • No code, outputs, per-repetition parameters or analysis notebooks are published.
  • D is defined as mixed but Tables 1 and 2 call it non-mixed.
  • The textual definition of neutral level contradicts the pseudocode.
  • The Claude Sonnet 2 header contradicts the Claude 3 Sonnet model.
  • The limitations contradict the appendix on which traits and models received BFI-44 validation.
  • Missing human intervals appear as 0.00/0.00 in the table instead of NA.
  • Differences between large and small models do not isolate size, architecture, provider, data or training.
  • Mini, Haiku and Flash do not constitute controlled distillation pairs in the study; the explanation by knowledge transfer is speculative.
  • The claim that Openness is the main factor depends on selecting alpha and beta and ignores the larger Agreeableness-lambda association.
  • Being close to CPT parameters equal to one in this prompt does not demonstrate general rationality.

What the study does not establish

  • It does not establish that LLMs possess internal or stable Big Five personality.
  • It does not demonstrate that a psychological cause called Openness produces risk.
  • It does not separate the trait effect from the direct meaning of the prompt adjectives.
  • It does not demonstrate stable preferences or choice under incentives.
  • It does not demonstrate general rationality of GPT-4o, Claude or Gemini.
  • It does not demonstrate equivalence between human cognition and LLM responses.
  • It does not demonstrate that distillation erases cognitive biases.
  • It does not demonstrate a universal hierarchy between advanced and small models.
  • It does not demonstrate behavior in markets, finance or deployed agents.
  • It does not allow exact reproduction of analyses and tests without code and raw data.

Traceability

Scope: Full text

Version: Findings of ACL 2025, Anthology ID 2025.findings-acl.1085, DOI 10.18653/v1/2025.findings-acl.1085, pages 21068-21092, 25 pages; supersedes arXiv:2503.04735v1

Consulted source: https://aclanthology.org/2025.findings-acl.1085/

Review: Codex complete bilingual full-text fidelity pass using the published ACL 2025 version, arXiv-to-publication reconciliation, all-page visual inspection, model/version and table extraction, CPT design audit, human-baseline comparability audit, statistical consistency review, lexical-confound analysis, reproducibility search, and internal cross-reference reconciliation; summaries written from the full paper and appendices rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • gpt-4o (2024-08-06)
  • gpt-4o mini (2024-07-18)
  • gpt-4 (turbo-2024-04-09)
  • claude-3-sonnet-20240229-v1.0
  • claude-3-haiku-20240307-v1.0
  • gemini-1.5-pro-002
  • gemini-1.5-flash-002

Instruments and metrics

  • Direct least-positive or most-negative certainty-equivalent prompt with step-by-step reasoning
  • Five-parameter cumulative prospect theory fit
  • IPIP-NEO-300
  • BFI-44 intervention check
  • Bipolar Big Five adjective persona prompts
  • Spearman correlation and t-statistic significance markers
  • 10,000-sample bootstrap confidence intervals with unspecified resampling unit

Data used

  • 56 mixed gain/loss prospects randomly sampled from choices13k; exact table printed, selection seed unreported
  • 56 non-mixed prospects and median certainty equivalents from 25 graduate students in Tversky and Kahneman (1992), used for historical comparison
  • Johnson (2020) IPIP-NEO-300 responses from 4,808 UK and Irish adults aged over 30
  • No released raw model outputs, fitted per-run parameters, or analysis code

Evidence and location

  • Version, authors, venue, DOI and pagination: ACL Anthology record and published PDF page 21068 checked 15 July 2026
  • Certainty equivalent method, CPT and choices13k: Sections 3.1-3.2, pages 21070-21072; Appendix C, pages 21086-21087
  • Models, versions, seeds and temperature: Appendix B.1-B.2, pages 21083-21084
  • Baseline CPT results: Figure 2 and Section 4.1, pages 21073-21074; Table 10, page 21090
  • Interventions and correlations by model and trait: Sections 4.2-4.3, Tables 1-2 and Figure 3, pages 21073-21076
  • Personality markers and level conflict: Appendix B.4, Figure 7 and Table 7, pages 21084-21085
  • Human data and sample sizes: Section 3.3 and Appendix C, pages 21072 and 21086; Table 9 and Section D.2, pages 21088-21089
  • BFI-44 validation and statistical legends: Appendix E, Tables 11-14, pages 21091-21092
  • Acknowledged limitations and internal inconsistencies: Limitations, page 21076, reconciled against Sections 3.2-4.3 and Appendices B-E
  • Complete audit of version, method and validity: reports/verification/article-187-cpt-personality-and-version-audit.json
  • Complete visual inspection: All 25 published PDF pages rendered and visually inspected on 15 July 2026