Persona-Conditioned Risk Behavior in Large Language Models: A Simulated Gambling Study with GPT-4.1

Applications, bias, and safety2026arXivApproved editorial review

Authors: Sankalp Dubedy

Keywords: Persona conditioning, Human simulation, Safety and bias

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

1
Authors
7
Findings
14
Limitations
5
Evidence

Editorial summary

English

The paper studies how three socioeconomic prompts alter GPT-4.1 behavior in a simulated slot machine. Rich starts with $10,000 and is instructed to preserve wealth and avoid unnecessary risk; Middle starts with $500 and should seek steady growth while managing risk; Poor starts with $50 and should take calculated risks to improve its situation. Each persona faces Fair (50% win), Biased Low (35%), and Streak machines (40% initially, +5 percentage points after each loss up to 80%). The study runs 50 sessions per combination, capped at 50 rounds: 450 sessions and 6,950 decisions. At each round the model returns PLAY/STOP, a bet, and several numeric and categorical self-assessments.

The three prompts produce very large separation. Rich plays 1.11 rounds on average, Middle 7.83, and Poor 37.39; session-length rank-biserial effects are 1.000 for Rich-Middle and Rich-Poor and 0.901 for Middle-Poor. Mean self-reported risk scores are 17.53, 40.23, and 63.36, respectively. The Fair machine receives a somewhat higher fairness score than Biased Low and Streak, although UNCERTAIN dominates categorical judgments. These data clearly show that different economic instructions produce persistent gambling policies in this GPT-4.1 configuration.

They do not, however, show that Prospect Theory emerges without instruction. The primary outcomes are embedded in the treatment: Rich is told to avoid risk and Poor to take it. Persona label, balance, goal, and reference framing also change together, with no neutral control or factorial design to isolate their effects. Prospect Theory would require controlled gain/loss choices relative to a reference point and estimation of loss aversion or probability weighting; this study does neither, and it has no human sample with which to validate 'human-like' behavior. Poor has a median of 50 rounds, exactly the cap, so at least half of its sessions are right-censored.

Most secondary analyses treat 6,950 rounds as independent even though they are nested within 450 sessions and 80.7% come from Poor because that condition keeps playing. Stopping creates informative selection: later rounds exist only for surviving sessions. Row-level ANOVA, correlations, and chi-square tests therefore give overly optimistic p-values without a multilevel model or session bootstrap. Correlating round number with 'risk score' is also an inadequate measure of belief updating: the score may express risk tolerance rather than machine probability, and a zero correlation can hide nonlinear learning, cancellation, or survivor selection.

Emotion, strategy, and decision labels are generated simultaneously in the same JSON. CAUTIOUS co-occurring with RISK_SEEKING does not establish that emotion is post-hoc or identify a causal direction. The broader warning against trusting self-explanations is reasonable, but this experiment provides suggestive inconsistency rather than a causal test. The study uses one model, one temperature, and one wording per persona; a fourth Explorer persona was removed by Azure's filter. Full prompts, code, data, seeds, message history, exact machine logic, and analysis are not released. The arXiv source contains only TeX and figures, and no study repository was located. The faithful conclusion is a strong demonstration of persistent prompt following, not a validated cognitive or socioeconomic replication.

Español

El artículo estudia cómo tres prompts socioeconómicos modifican la conducta de GPT-4.1 en una tragaperras simulada. La persona Rich empieza con 10.000 dólares y recibe la instrucción de preservar riqueza y evitar riesgo innecesario; Middle empieza con 500 y debe crecer de forma estable gestionando el riesgo; Poor empieza con 50 y debe asumir riesgos calculados para mejorar su situación. Cada persona afronta máquinas Fair (50% de victoria), Biased Low (35%) y Streak (40% inicial, +5 puntos tras cada pérdida hasta 80%). Se ejecutan 50 sesiones por combinación, con un máximo de 50 rondas: 450 sesiones y 6.950 decisiones. En cada ronda el modelo devuelve PLAY/STOP, apuesta y varias autoevaluaciones numéricas y categóricas.

Los tres prompts generan una separación muy grande. Rich juega una media de 1,11 rondas, Middle 7,83 y Poor 37,39; las comparaciones de duración tienen efectos rank-biserial de 1,000 para Rich-Middle y Rich-Poor y 0,901 para Middle-Poor. Los riesgos autodeclarados medios son 17,53, 40,23 y 63,36, respectivamente. La máquina Fair recibe una puntuación de justicia algo mayor que Biased Low y Streak, aunque UNCERTAIN domina los juicios. Estos datos muestran con claridad que instrucciones económicas distintas producen políticas de juego persistentes en esta configuración de GPT-4.1.

No prueban, sin embargo, que Prospect Theory emerja sin instrucción. Los resultados principales están contenidos en el tratamiento: Rich debe evitar riesgo y Poor debe asumirlo. Además cambian simultáneamente etiqueta, saldo, objetivo y referencia, sin control neutral ni diseño factorial que permita separar sus efectos. Prospect Theory requeriría comparar ganancias y pérdidas controladas respecto a un punto de referencia y estimar aversión a pérdidas o ponderación de probabilidades; aquí no se hace, ni existe una muestra humana que valide que la conducta sea «human-like». La mediana de Poor es 50, exactamente el límite, por lo que al menos la mitad de sus sesiones está censurada por techo.

La mayoría de análisis secundarios usa las 6.950 rondas como si fueran independientes, aunque están anidadas en 450 sesiones y el 80,7% procede de Poor porque esa condición sigue jugando. El cese produce selección informativa: las rondas tardías solo existen para sesiones supervivientes. Por ello, ANOVA, correlaciones y chi-cuadrado a nivel de fila ofrecen p-valores demasiado optimistas sin un modelo multinivel o bootstrap por sesión. La correlación entre número de ronda y «risk score» tampoco mide adecuadamente actualización de creencias: el score puede expresar tolerancia al riesgo, no probabilidad de la máquina, y una correlación nula puede ocultar aprendizaje no lineal, cancelación o selección.

Las emociones, estrategias y decisiones se generan simultáneamente en el mismo JSON. Que CAUTIOUS coexista con RISK_SEEKING no establece que la emoción sea una narración posterior ni identifica causalidad. La idea general de no confiar en autoexplicaciones es razonable, pero el experimento solo aporta una inconsistencia sugestiva. El estudio se limita a un modelo, una temperatura y una redacción por persona; una cuarta persona Explorer fue excluida por el filtro de Azure. No se publican prompts completos, código, datos, semillas, historial de mensajes, lógica exacta de la máquina ni análisis. La fuente arXiv contiene solo TeX y figuras y no se localizó un repositorio. La conclusión fiel es una demostración fuerte de seguimiento persistente de prompts, no una replicación cognitiva o socioeconómica validada.

Research question

How does the sequential behavior of GPT-4.1 change in a slot machine when assigned three economic identities with different balances and risk objectives, and what relationship do its self-labels of risk, emotion, strategy, and fairness bear to the decisions and accumulated experience?

Method

3x3 design with Rich, Middle, and Poor personas and Fair, Biased Low, and Streak machines. Each cell contains 50 independent sessions according to the author and each session has a maximum of 50 rounds. GPT-4.1 is used through Azure OpenAI with temperature 1.0 and structured JSON. The responses include PLAY/STOP, bet, risk/confidence/fairness/reward expectation/uncertainty, emotion, strategy, and reasoning. 450 sessions and 6,950 decisions are analyzed with Kruskal-Wallis, Mann-Whitney, ANOVA, Cohen d, point-biserial and Spearman correlations, chi-square, and Cramér's V. Bonferroni is applied only to three primary duration comparisons.

Sample: 450 sessions: 3 personas x 3 machines x 50 iterations, with a maximum of 50 rounds. The personas contribute 166 Rich decisions, 1,175 Middle, and 5,609 Poor; therefore, Poor represents 80.7% of the rows. There are no human participants.

Findings

  • Rich, Middle, and Poor show 1.11, 7.83, and 37.39 mean rounds; the three prompts produce extreme and consistent separation.
  • The median for Poor reaches the maximum of 50 rounds, so at least half of those sessions is censored by ceiling.
  • The mean self-declared risk scores follow directly the order induced by the objectives: 17.53, 40.23, and 63.36.
  • Fair receives a fairness score of 59.99 compared to 54.27 for Biased Low and 55.49 for Streak, but UNCERTAIN dominates the three categorical judgments.
  • The main effect demonstrates persistent adherence to instructions, not uninstructed emergence of Prospect Theory.
  • The emotion and belief analyses are associations between simultaneous self-labels and do not identify causality or learning.
  • The arXiv source provides TeX and 13 figures, but no code, complete prompts, data, logs, or analysis scripts.

Limitations

  • The instructions prescribe avoiding or assuming risk, creating circularity between treatment and outcome.
  • Socioeconomic label, balance, objective, and reference change together; there is no neutral control or factorial separation.
  • Equivalent gains and losses are not manipulated, nor is the value or weighting function of Prospect Theory estimated.
  • There is no human comparison to validate realism, socioeconomic bias, or human-like behavior.
  • A single wording represents each construct; repetitions do not substitute prompt variants.
  • Poor is censored by the 50-round limit and Rich by the floor of one round.
  • Per-round tests ignore within-session and condition dependence, and stopping induces informative selection.
  • 80.7% of decisions come from Poor, dominating aggregate statistics.
  • Risk score is an ambiguous self-label and not a calibrated probability estimate.
  • Emotion, strategy, and decision are produced simultaneously; chi-square does not establish temporal or causal order.
  • Only GPT-4.1, temperature 1.0, and an Azure deployment are used, with no immutable snapshot.
  • Explorer is excluded by moderation without publishing the prompt, filter category, or correction attempts.
  • RNG, seeds, retries, JSON errors, betting rules, bankroll, or Streak reset are not specified.
  • No code, data, sessions, complete prompts, or reproducible analysis are published.

What the study does not establish

  • It does not demonstrate that GPT-4.1 reproduces Prospect Theory without risk instructions.
  • It does not demonstrate human behavior or real differences between rich, middle, and poor personas.
  • It does not separately identify the effect of wealth, label, balance, objective, or wording.
  • It does not demonstrate that emotional self-labels are subsequent to or cause decisions.
  • It does not demonstrate absence of belief updating through a pooled correlation with round number.
  • It does not allow inferences from 6,950 independent units because decisions are nested within 450 sessions.
  • It does not estimate the natural stopping time of Poor due to censoring at 50 rounds.
  • It does not generalize to other models, temperatures, prompts, economic tasks, or decisions with real consequences.
  • It does not allow reproducing results without experimental artifacts and a versioned model.

Traceability

Scope: Full text

Version: arXiv:2603.15831v1, submitted 2026-03-16, CC BY 4.0

Consulted source: https://arxiv.org/abs/2603.15831

Review: Codex 20-page visual full-text, prompt-confounding, Prospect-Theory construct, censoring, nested-statistics, self-report causality, learning and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1 via Azure OpenAI

Instruments and metrics

  • Simulated sequential slot-machine task
  • PLAY/STOP and bet amount
  • Model self-reported risk, confidence, fairness, reward expectation and uncertainty scores
  • Model self-reported emotion, strategy and fairness labels
  • Kruskal-Wallis and Mann-Whitney U
  • One-way ANOVA and Cohen's d
  • Point-biserial and Spearman correlations
  • Chi-square and Cramer's V

Data used

  • 6,950 unreleased GPT-4.1 decision records from 450 simulated sessions

Evidence and location

  • Design, summarized prompts, model, sample, machines, results, discussion, limitations, and SBI proposal: arXiv:2603.15831v1, 20/20 PDF pages rendered and individually inspected
  • Version v1, date, author, license, 21-page comment, and absence of links to artifacts: Official arXiv abstract and Atom records inspected 2026-07-17
  • The official source only contains TeX, figures, and metadata, with no code or data: Official arXiv e-print source archive SHA-256 0d1e2e5b549287595ddddda0b157dde3cccc1ec77a85f8516e1f74e82f4990cc inspected 2026-07-17
  • No current location of the study repository: Exact-title, arXiv-ID, author, web and GitHub repository searches inspected 2026-07-17
  • Audit of prompt confounding, Prospect Theory, censoring, pseudoreplication, self-labels, causality, learning, and reproducibility: reports/verification/article-390-gpt41-gambling-prompt-confounding-prospect-theory-pseudoreplication-ceiling-self-report-causality-and-reproducibility-audit.json