Investigating the Impact of LLM Personality on Cognitive Bias Manifestation in Automated Decision-Making Tasks

Applications, bias, and safety2025arXivApproved editorial review

Original title: Investigating the Impact of LLM Personality on the Manifestation of Cognitive Biases in Decision-Making Tasks

Authors: Jiangen He, Jiqun Liu

Keywords: Artificial Intelligence, Large Language Models, Cognitive Biases, Decision-Making, Personality Traits

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

2
Authors
20
Findings
34
Limitations
10
Evidence

Editorial summary

English

The preprint asks whether instructions describing the Big Five traits change eight decision effects in GPT-4o, GPT-4o-mini, Llama 3-70B, and Llama 3-8B, and whether a verbal warning reduces those effects. Personality needs precise interpretation here: the study neither measures an inherent model trait nor validates that a model adopts a psychometric profile. It prepends a description of Openness, Conscientiousness, Extraversion, Agreeableness, or Neuroticism, or a supposedly reversed description. The material combines 13,465 university-admission prompts reused from Echterhoff et al. for anchoring, framing, status quo, and group attribution with 4,585 GPT-4-generated BiasEval scenarios for decoy, risk aversion, sunk cost, and endowment effects. All four models run at temperature zero under baseline, trait, and reversed-trait conditions; mitigation literally adds the sentence, “Be mindful of not being biased by cognitive bias.” Each effect is a difference in rates except endowment, which normalizes the willingness-to-accept minus willingness-to-pay difference by a control valuation. Mitigation is the reduction in absolute effect relative to baseline, so a negative mitigation value means that magnitude increased. Results depend strongly on architecture and prompt. For anchoring, GPT-4o moves from 0.112 without a trait to 0.338 under Agreeableness, while Llama 3-70B starts at -0.243 and remains negative under all five traits. For framing, GPT-4o reduces a -0.063 baseline difference to between -0.008 and 0.002, but Llama 3-70B increases the magnitude to -0.181 under Neuroticism. Decoy baselines range from -0.385 for GPT-4o-mini to 0.128 for Llama 3-70B; a negative value is a shift opposite the predicted decoy direction, not absence of sensitivity. Risk-aversion baselines are 0.044, 0.544, 0.428, and 0.074 for GPT-4o, GPT-4o-mini, Llama 3-70B, and Llama 3-8B, with trait prompts sometimes increasing and sometimes decreasing them. Sunk cost is almost always zero, except 0.008 and 0.019 for Extraversion and Agreeableness in GPT-4o. Status-quo baselines are 0.134, -0.101, -0.320, and -0.004, again with heterogeneous directions. Endowment changes are extreme: published baselines are 4.27%, 23.07%, 91.23%, and 39.77%, while some trait conditions reach 109.59% for GPT-4o and 124.37% for Llama 3-8B. Baseline gender group-attribution effects are small (-0.002, -0.018, 0.019, and 0). The paper concludes that Conscientiousness and Agreeableness tend to help the awareness warning. Recomputing its own descriptive figure gives mean reductions of 0.0305 for Conscientiousness and 0.0246 for Agreeableness over 24 rows each, but the figure duplicates every framing value under status quo for all four models. No tests, intervals, standard errors, or independent replications justify using “significantly” in the statistical sense. Reproducibility is also inadequate: although the appendix promises detailed prompts, it leaves `{context}`, `{question}`, and other placeholders instead of the actual personality, reversed-personality, and scenario text; BiasEval, generations, code, model snapshots, and API dates are not released. The summary table omits Openness for GPT-4o-mini anchoring; the detailed table provides neither a baseline nor mitigation values for Llama 3-8B anchoring even though the main table reports them; and the detailed sunk-cost table omits GPT-4o-mini. The defensible contribution is evidence that outputs on synthetic decision tasks change substantially with instructions and architecture. It does not establish stable model traits, psychological causation, general efficacy of any personality for debiasing, or safety in real decisions.

Español

El preprint estudia si instrucciones que describen los cinco grandes rasgos de personalidad modifican ocho efectos de decisión en GPT-4o, GPT-4o-mini, Llama 3-70B y Llama 3-8B, y si una advertencia verbal reduce esos efectos. Conviene precisar desde el principio qué significa aquí personalidad: no se mide un rasgo inherente del modelo ni se valida que este adopte un perfil psicométrico; se antepone una descripción de apertura, responsabilidad, extraversión, amabilidad o neuroticismo, o una descripción supuestamente inversa. El material combina 13.465 prompts de admisión universitaria reutilizados de Echterhoff et al. para anclaje, framing, statu quo y atribución grupal con 4.585 escenarios BiasEval generados mediante GPT-4 para efecto señuelo, aversión al riesgo, coste hundido y efecto dotación. Los cuatro modelos se ejecutan con temperatura 0 bajo condiciones base, de rasgo y de rasgo invertido; para la mitigación se añade literalmente “Be mindful of not being biased by cognitive bias”. Cada efecto se expresa como una diferencia de tasas, salvo dotación, que normaliza la diferencia entre willingness-to-accept y willingness-to-pay con una valoración de control. La mitigación se calcula como la reducción del valor absoluto del efecto respecto de la condición base, de modo que un signo negativo en la columna de mitigación significa aumento de magnitud. Los resultados son muy dependientes del modelo y del prompt. En anclaje, GPT-4o pasa de 0,112 sin rasgo a 0,338 con amabilidad, mientras Llama 3-70B parte de -0,243 y permanece negativo en los cinco rasgos. En framing, GPT-4o reduce una diferencia base de -0,063 a valores entre -0,008 y 0,002, pero Llama 3-70B aumenta la magnitud hasta -0,181 con neuroticismo. Los baselines del señuelo varían de -0,385 en GPT-4o-mini a 0,128 en Llama 3-70B; un valor negativo indica desplazamiento en dirección contraria al efecto esperado, no ausencia de sensibilidad. En aversión al riesgo los baselines son 0,044, 0,544, 0,428 y 0,074 para GPT-4o, GPT-4o-mini, Llama 3-70B y Llama 3-8B; las instrucciones de rasgo pueden tanto ampliar como reducir esas diferencias. Coste hundido queda en cero casi siempre, salvo 0,008 y 0,019 para extraversión y amabilidad en GPT-4o. Los baselines de statu quo son 0,134, -0,101, -0,320 y -0,004, de nuevo con direcciones heterogéneas. El efecto dotación muestra cambios extremos: los baselines publicados son 4,27%, 23,07%, 91,23% y 39,77%, pero con algunos rasgos alcanzan 109,59% en GPT-4o y 124,37% en Llama 3-8B. La atribución grupal por género es pequeña en las condiciones base (-0,002, -0,018, 0,019 y 0). El paper resume que responsabilidad y amabilidad suelen favorecer la advertencia de mitigación. Recalculando su propia figura descriptiva, la reducción media es 0,0305 para responsabilidad y 0,0246 para amabilidad en 24 filas cada una, pero la figura duplica exactamente todos los valores de framing bajo la etiqueta statu quo en los cuatro modelos. No hay tests, intervalos, errores estándar ni repeticiones independientes que permitan usar “significativamente” en sentido estadístico. La reproducibilidad también es insuficiente: el apéndice promete prompts detallados, pero deja `{context}`, `{question}` y otras variables sin las descripciones reales de personalidad, personalidad inversa o escenarios; no se publican BiasEval, salidas, código, snapshots de modelos ni fechas de API. La tabla resumida omite apertura para GPT-4o-mini en anclaje, la tabla detallada deja sin baseline ni mitigación las filas de anclaje de Llama 3-8B aunque la tabla principal sí da esos valores, y la tabla detallada de coste hundido omite GPT-4o-mini. La contribución defendible es demostrar que las respuestas a tareas sintéticas cambian mucho al modificar instrucciones y arquitectura. No demuestra rasgos estables del modelo, causalidad psicológica, eficacia general de una personalidad para mitigar sesgos ni seguridad en decisiones reales.

Research question

How do eight measures of bias or decision effect change when four LLMs are instructed to represent normal or inverted Big Five traits, and to what extent does a zero-shot warning not to be biased reduce the magnitude of those effects?

Method

13,465 synthetic admission prompts for anchoring, framing, status quo, and group attribution are reused, and another 4,585 BiasEval scenarios for decoy, risk aversion, sunk cost, and endowment are generated with GPT-4. GPT-4o, GPT-4o-mini, Llama 3-70B, and Llama 3-8B are queried at temperature 0 under no-trait conditions, with each of the five traits, and with inverted traits; mitigation adds a generic warning. Differences in rates or ratings between pairs of conditions are compared and mitigation is defined as the reduction of the absolute value relative to the baseline. The editorial audit read and rendered the 19 pages, froze and reviewed the LaTeX source, checked the tables and the figure, recalculated its descriptive aggregates, and searched for associated code and data.

Sample: 18,050 declared base scenarios or prompts: 13,465 admission and 4,585 BiasEval. The admission corpus includes 5,449 anchoring prompts, 1,008 status quo prompts duplicated for control, 1,000 framing prompts triplicated, and 1,000 group attribution prompts triplicated by gender. BiasEval contains between 1,000 and 1,300 scenarios for each of four effects. These materials are reused across four models, five traits, inverted conditions, baseline, and warning, but the paper does not report the total number of calls, failures, exclusions, or repetitions.

Findings

  • The identified current source is arXiv:2502.14219v1; the author's current page still classifies it as a preprint.
  • The study combines 13,465 admission prompts with 4,585 unpublished BiasEval scenarios.
  • The four models run at temperature 0, but no snapshots, dates, or full configurations are identified.
  • Personality means a trait instruction; no resulting Big Five profile is measured or validated.
  • The baseline anchoring is 0.112 in GPT-4o and rises to 0.338 with agreeableness.
  • Llama 3-70B has baseline anchoring -0.243 and negative values between -0.202 and -0.260 depending on the trait.
  • GPT-4o reduces framing from -0.063 at baseline to a range of -0.008 to 0.002 under traits.
  • Llama 3-70B increases framing from -0.052 at baseline up to -0.181 with neuroticism.
  • The decoy effect baselines are 0.036, -0.385, 0.128, and 0.103 for GPT-4o, GPT-4o-mini, Llama 3-70B, and Llama 3-8B.
  • The risk aversion baselines are 0.044, 0.544, 0.428, and 0.074 in the same order.
  • Sunk cost is zero in almost all cells; only GPT-4o shows 0.008 with extraversion and 0.019 with agreeableness.
  • The status quo baselines are 0.134, -0.101, -0.320, and -0.004, which shows different directions across models.
  • The baseline endowment effects are 4.27%, 23.07%, 91.23%, and 39.77%, with trait values exceeding 100%.
  • The baseline group attribution by gender is small: -0.002, -0.018, 0.019, and 0.
  • Inverted traits do not systematically produce the opposite effect to their normal versions.
  • Mitigation is limited to adding an awareness sentence and comparing absolute magnitudes.
  • In the figure, conscientiousness and agreeableness have the largest mean descriptive reductions: 0.0305 and 0.0246.
  • The 24 rows of the figure include four exact copies: status quo repeats framing for each model.
  • No statistical tests, intervals, errors, or replications supporting inferential significance are reported.
  • No repository, BiasEval, outputs, execution code, or associated public analysis was located.

Limitations

  • Traits are induced by prompt and not observed inherent properties of the model.
  • No Big Five inventory is administered nor is it checked that each condition adopts the intended trait.
  • The exact personality instructions do not appear in the appendix or in the source.
  • The exact inverted personality instructions are not published either.
  • The appendix replaces scenarios with markers such as `{context}`, `{question}`, and `{condition}`.
  • BiasEval was generated with GPT-4 but is not released or validated with human judges.
  • No seed, generation prompt, or filtering process for BiasEval is reported.
  • The admission and BiasEval scenarios are synthetic and do not represent deployed decisions.
  • Only four models from a specific time window are studied.
  • Exact IDs for GPT-4o and GPT-4o-mini and query dates are missing.
  • Checkpoints, revisions, quantization, server, and chat templates for Llama 3 are missing.
  • Temperature 0 does not guarantee determinism across services or hardware and no replications are reported.
  • The total number of calls, errors, retries, invalid responses, or exclusions is not reported.
  • There are no hypothesis tests, confidence intervals, standard errors, or power analyses.
  • The use of `significantly` is descriptive, not inferential.
  • Negative values are interpreted as inverse bias, but their substantive importance is not validated.
  • Absolute-value-based mitigation equates a reduction toward zero with changes that may cross direction.
  • The warning literally mentions cognitive bias without specifying which one or controlling experimental demand.
  • There is no length or tone equivalent instruction control without mitigation content.
  • The effect of the trait content is not separated from obeying an additional instruction.
  • The endowment metric depends on generated numerical values without cleaning, dispersion, or robustness to extremes.
  • The relative mitigation of endowment uses baselines near zero and produces unstable percentages of up to -2467.85%.
  • The figure exactly duplicates framing as status quo in the four models.
  • Duplicating those rows alters the averages used to support the conclusion about traits.
  • The main table omits openness for GPT-4o-mini in anchoring, although the detailed table does include it.
  • The detailed anchoring table does not give baseline or mitigation for Llama 3-8B, but the main table does publish those numbers.
  • The detailed sunk cost table omits GPT-4o-mini even though the summary table shows zeros.
  • The main table publishes inverted anchoring mitigations for Llama 3-8B without underlying rates to verify them.
  • There is no code to reconstruct tables, colors, figure, or aggregates.
  • There are no output data to audit parsing, out-of-format responses, or sensitivity to specific examples.
  • Accuracy, utility, fairness for real groups, and consequences of decisions are not evaluated.
  • There is no human evaluation of whether the responses are reasonable or harmful.
  • Comparing patterns with human literature does not turn instructions into psychological personality.
  • The version is a preprint and no accepted peer-reviewed publication was identified.

What the study does not establish

  • It does not demonstrate that the models possess stable or inherent Big Five traits.
  • It does not prove that a trait instruction produces the same construct measured in humans.
  • It does not establish psychological causal relationships between personality and bias.
  • It does not demonstrate statistical significance of the observed differences.
  • It does not demonstrate that conscientiousness or agreeableness mitigate biases in a general way.
  • It does not validate the inverted conditions as psychometric opposites.
  • It does not allow reproducing the results without prompts, BiasEval, outputs, and snapshots.
  • It does not generalize from synthetic scenarios to real admission, health, finance, or hiring.
  • It does not establish that reducing a rate difference necessarily improves fairness or quality.
  • It does not demonstrate safety, absence of discrimination, or operational efficacy of the warning.

Traceability

Scope: Full text

Version: arXiv:2502.14219v1, submitted 20 February 2025; 19 pages

Consulted source: https://arxiv.org/pdf/2502.14219v1

Review: Codex full-text, visual, LaTeX-source and statistical-reporting audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o, exact snapshot and API date not reported
  • GPT-4o-mini, exact snapshot and API date not reported
  • Llama 3-70B, exact checkpoint and inference stack not reported
  • Llama 3-8B, exact checkpoint and inference stack not reported
  • GPT-4 used to generate the unreleased BiasEval scenarios, exact snapshot not reported

Instruments and metrics

  • Big Five trait descriptions used as prompt prefixes
  • Reversed-personality prompt conditions
  • Awareness instruction: Be mindful of not being biased by cognitive bias
  • Anchoring effect as admit-rate difference after a prior admit versus reject decision
  • Framing effect as admit-rate difference under admit versus reject wording
  • Decoy effect as target-option selection shift after adding a dominated option
  • Risk-aversion effect as risky-choice difference across gain and loss framing
  • Sunk-cost effect as continuation-rate difference with versus without past investment
  • Status-quo effect as default-option rate minus mean alternative-option rate
  • Endowment effect from WTA, WTP and control valuations
  • Gender group-attribution effect from male versus female math judgments

Data used

  • Student Admission Dataset from Echterhoff et al. (2024), 13,465 prompts
  • Unreleased BiasEval dataset, 4,585 GPT-4-generated scenarios
  • Published summary and appendix tables for eight effects
  • Published bias-reduction figure with duplicated framing/status-quo rows
  • arXiv:2502.14219v1 LaTeX source archive

Evidence and location

  • Question, framework, and definition of personality: arXiv v1 pp. 1–3, Abstract, Sections 1 and 2.1
  • Eight effects and taxonomy: arXiv v1 pp. 3–5, Sections 2.2–2.4
  • Admission corpus and BiasEval: arXiv v1 p. 4, Sections 2.3.1–2.3.2
  • Models, temperature, and mitigation metric: arXiv v1 p. 5, Section 2.4 and opening of Section 3
  • Results for normal and inverted traits: arXiv v1 pp. 5–8, Tables 1–3 and Sections 3.1–3.2
  • Warning and mitigation conclusion: arXiv v1 pp. 7–8, Section 3.3 and Figure 2
  • Limitations acknowledged by the authors: arXiv v1 p. 9, Section 6
  • Incomplete prompts and detailed tables: arXiv v1 pp. 11–19, Appendix A and Tables 4–11
  • Framing/status quo duplication and aggregates by trait: Frozen arXiv v1 source asset bias-reduction.pdf, independently extracted and recomputed 15 Jul 2026
  • Preprint status and absence of public artifacts: arXiv record, current author publication page, Papers With Code and GitHub title/author search, checked 15 Jul 2026