The Impact of Big Five Personality Traits on AI Agent Decision-Making in Public Spaces: A Social Simulation Study

Society, culture, and collective behavior2025arXivApproved editorial review

Authors: Minjun Ren, Wentao Xu

Keywords: Big Five personality, AI agent decision-making, social simulation, multi-agent framework, public spaces, AgentVerse, behavioral modeling

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

2
Authors
5
Findings
7
Limitations
7
Evidence

Editorial summary

English

This preprint examines whether ten GPT-3.5-turbo agents, defined through opposing Big Five labels, produce different patterns when judging misinformation in an AgentVerse classroom simulation. Each pair represents the poles of one domain: curious/cautious, organized/careless, outgoing/reserved, friendly/critical, and sensitive/confident. A facilitator agent acts as professor. Each student presents a different false claim, and the other nine generate a public opinion marked [Speak] and a supposedly private text marked [Think]; silences are excluded. Before and after the simulation, agents rate adapted personality statements on a 1–5 scale to check whether their persona instructions remain visible. The stability table appears to summarize 50 runs, but the paper does not reproducibly specify the number of iterations, full prompts, generation parameters, or exact GPT-3.5-turbo snapshot. Descriptive counts show the largest contrast between the curious agent, with 100 public yes and 8 no responses, 92.6% acceptance, and the cautious agent, with 4 yes and 223 no responses, 97.8% rejection. The critical agent also rejects frequently, 23 yes and 207 no, while the remaining labels are more balanced. [Speak]–[Think] discrepancies are largest for friendly (158), outgoing (132), careless (130), and sensitive (106), and smallest for cautious (6), confident (15), critical (31), and curious (35). The manuscript interprets these patterns as effects of openness, conscientiousness, and extraversion on information acceptance and as social sensitivity in some profiles. This interpretation is exploratory. No correlation coefficients, significance tests, confidence intervals, or statistical model are reported, even though the abstract refers to “significant correlations.” There is also no unlabelled control condition, prompt randomization, replication across models, or human evaluation. [Think] is not access to a private belief: it is another requested text output in the same interaction, so disagreement with [Speak] does not demonstrate cognitive dissonance. The ten topics, interaction order, role of the misinformation presenter, and agent label can be confounded with the claimed trait effect. Binary adjectives stand in for trait poles rather than a validated Big Five instrument or continuous scores. The study therefore documents different output frequencies from one GPT-3.5 configuration under ten instructed personas; it does not establish psychological traits, decision mechanisms, or externally valid human-behavior simulation. Several generic references also lack persistent identifiers and were not found in the primary publication indexes checked, weakening the theoretical audit trail.

Español

Este preprint estudia si diez agentes basados en GPT-3.5-turbo, definidos mediante etiquetas opuestas de los Big Five, producen patrones diferentes al valorar desinformación en una simulación de aula con AgentVerse. Cada par representa los extremos de un dominio: curioso/cauteloso, organizado/descuidado, extrovertido/reservado, amistoso/crítico y sensible/confiado. Un agente facilitador actúa como profesor. Cada estudiante presenta una afirmación falsa distinta y los otros nueve generan una opinión pública marcada como [Speak] y un texto supuestamente privado marcado como [Think]; los silencios se excluyen. Antes y después de la simulación, los agentes puntúan en una escala 1–5 enunciados adaptados de un estudio anterior para comprobar que sus instrucciones de personalidad siguen visibles. La tabla de estabilidad parece resumir 50 ejecuciones, aunque el artículo no describe de forma reproducible el número de iteraciones, los prompts completos, los parámetros de generación ni la instantánea exacta de GPT-3.5-turbo. Los conteos descriptivos muestran el contraste mayor entre el agente curioso, con 100 respuestas públicas afirmativas y 8 negativas, 92,6 % de aceptación, y el cauteloso, con 4 afirmativas y 223 negativas, 97,8 % de rechazo. El agente crítico también rechaza con frecuencia, 23 sí y 207 no; los resultados de las demás etiquetas son más equilibrados. La discrepancia [Speak]–[Think] es mayor para amistoso (158 casos), extrovertido (132), descuidado (130) y sensible (106), y menor para cauteloso (6), confiado (15), crítico (31) y curioso (35). El manuscrito interpreta estos patrones como influencia de apertura, responsabilidad y extraversión sobre la aceptación de información y como sensibilidad social en algunos perfiles. Esa lectura debe tomarse como exploratoria. No se reportan coeficientes de correlación, pruebas de significación, intervalos de confianza ni un modelo estadístico, pese a que el abstract habla de «correlaciones significativas». Tampoco existe una condición sin etiquetas de personalidad, aleatorización de prompts, replicación con otros modelos o evaluación humana. [Think] no es acceso a una creencia privada del sistema: es otra salida textual solicitada en el mismo prompt, por lo que su diferencia con [Speak] no demuestra cognición disonante. Los diez temas, el orden de interacción, el rol de quien presenta cada falsedad y la etiqueta del agente pueden confundirse con el efecto atribuido al rasgo. Los extremos se expresan mediante adjetivos binarios, no mediante un instrumento Big Five validado ni puntuaciones continuas. En consecuencia, el estudio documenta frecuencias distintas de texto generado por una única configuración de GPT-3.5 bajo diez personas instruidas; no establece rasgos psicológicos, mecanismos de decisión ni comportamiento humano simulado con validez externa. Además, varias referencias genéricas del manuscrito carecen de identificadores y no se localizaron en los índices editoriales primarios consultados, lo que reduce la trazabilidad de su marco teórico.

Research question

Which opposing Big Five labels most clearly modify the public responses and [Think] texts of GPT-3.5 agents when faced with misinformation in a classroom social simulation?

Method

Descriptive multiagent simulation with AgentVerse. Ten instances of GPT-3.5-turbo receive pairs of opposite adjectives representing the five Big Five domains and an eleventh agent moderates. In classroom rounds, each student presents one of ten false statements and the others respond with [Speak] and [Think] segments. Affirmative responses, negative responses, silences, and discrepancies between both segments are counted. A pretest and posttest of personality statements on a 1–5 scale are used as a control of prompt persistence; the article does not report inferential tests or human evaluation.

Sample: Ten persona-agents and one facilitator, all based on GPT-3.5-turbo. The pre/post table contains 50 comparisons per person; the behavioral table retains between 108 and 330 responses per person after excluding silences. There are no human participants, annotators, or decisions.

Findings

  • The curious agent accepts 100 of 108 statements publicly (92.6 %), while the cautious one rejects 223 of 227 (97.8 %), the broadest descriptive contrast.
  • The critical profile rejects 207 of 230 statements (90 %); responsibility and extraversion also show frequency contrasts, but smaller and without statistical estimation of effect.
  • The largest discrepancies between [Speak] and [Think] appear in friendly (158), extroverted (132), careless (130), and sensitive (106); the smallest in cautious (6) and confident (15).
  • The pretest and posttest means are close for most profiles, although friendly has a mean difference of 0.855 and only 34 % of cases below 0.5; therefore stability is not equally strong across all agents.
  • The article provides counts and percentages, not the correlations or significance tests that its wording claims.

Limitations

  • A single closed family and an unspecified version prevent separating a personality effect from peculiarities of GPT-3.5-turbo and from subsequent service changes.
  • Complete prompts, sampling parameters, seeds, logs, specific code, and an unambiguous description of the iterations are not published, so the experiment is not reproducible with the article.
  • Binary labels of two adjectives do not constitute a psychometrically validated manipulation of the Big Five, and the pre/post measures compliance with the same instructed character.
  • There is no neutral condition, counterbalancing, randomization, replication across models, or statistical tests; the expressions "significant" and "correlation" are not supported by reported analyses.
  • Each false topic is assigned to a specific agent and the interaction is sequential, which mixes personality, content, role, and social context.
  • [Think] is text generated and observable by the experimental system, not an independent measurement of private thought or belief; calling its difference "cognitive dissonance" anthropomorphizes the protocol.
  • Several references in the theoretical framework lack persistent identifiers and do not appear in the primary indexes consulted; they should be verified before reuse.

What the study does not establish

  • It does not demonstrate that the agents have personality, private beliefs, cognition, dissonance, or internal decisions comparable to human ones.
  • It does not demonstrate statistically significant correlations between Big Five and behavior; it only presents descriptive frequencies under persona prompts.
  • It does not test that openness is in general the most influential trait or that the results generalize to other models, topics, languages, environments, or formulations.
  • It does not allow interpreting [Speak] as real public behavior or [Think] as a mental state inaccessible to other agents.
  • It does not validate AgentVerse as a model of a human classroom or responses to misinformation as predictors of real social behavior.
  • It does not establish that personalizing agents through these labels improves safety, truthfulness, or user experience.

Traceability

Scope: Full text

Version: arXiv:2503.15497v1

Consulted source: https://arxiv.org/pdf/2503.15497

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • GPT-3.5-turbo (snapshot not specified)

Instruments and metrics

  • Ten binary adjective persona descriptions mapped to Big Five domains
  • Adapted pre/post personality statements with a five-point Likert scale
  • Binary [Speak] and [Think] response coding

Data used

  • Ten author-selected misinformation statements
  • AgentVerse-generated classroom interaction logs

Evidence and location

  • Objectives, model, and main claims: arXiv v1, p. 1, Abstract and Introduction
  • Agent design, labels, and AgentVerse environment: arXiv v1, pp. 2–3, Methodology and Table 1
  • Misinformation topics and [Speak]/[Think] protocol: arXiv v1, p. 3, Public Space Simulation and Table 2
  • Pre/post and limits of the alleged stability: arXiv v1, p. 3, Personality Consistency Test and Table 3
  • Frequencies per person and discrepancies: arXiv v1, p. 4, Figure 2 and Table 4
  • Causal interpretation proposed by the authors: arXiv v1, pp. 4–5, Discussion
  • Traceability of the theoretical framework: arXiv v1, pp. 1–2 and 5, Related Work and References