Can LLMs Truly Embody Human Personality? Analyzing AI and Human Behavior Alignment in Dispute Resolution

Personas, identity, and agents2026arXivApproved editorial review

Authors: Deuksin Kwon, Kaleen Shrestha, Bin Han, Spencer Lin, James Hale, Jonathan Gratch, Maja Matarić, Gale M. Lucas

Keywords: Large Language Models, Personality, Big Five, Persona, Personality Control

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

8
Authors
7
Findings
34
Limitations
13
Evidence

Editorial summary

English

This paper asks whether Big Five-prompted LLM agents reproduce the personality-behavior relationships observed in people during conflict. It compares a reported subset of 248 human-human KODIS dialogues with 1,000 same-model LLM negotiations: 500 GPT-4o-mini and 250 each for Claude 3.7 Sonnet and Gemini 2.0 Flash. All negotiate the same buyer-seller jersey dispute across five refund, review, and apology decisions. Final outcomes are payoff, acceptance, and not walking away; strategies are classified under the Interests-Rights-Power framework as cooperative, neutral, competitive, or residual, with additional reciprocity, escalation, and de-escalation metrics.

Agents receive 15 adjectives, three per trait, selected from 70 bipolar pairs and modified to encode six levels. Their marginal trait distributions imitate the human sample, but measurement is not equivalent: people answer ten BFI items whereas LLMs receive discretized labels in a prompt. The design also inserts a human association in advance: apology importance is made dependent on agreeableness using a KODIS regression (B=2.13, p=.02), while other issue weights are random. Any agreeableness effect on outcomes therefore combines the proposed psychological mechanism with an author-programmed utility preference.

In KODIS, score has no personality effects; acceptance is negatively associated with self-neuroticism (B=-.26, p=.026) and positively with partner neuroticism (B=.27, p=.025); not walking away has none. LLMs produce different sets of significant coefficients across models and outcomes. Humans rely heavily on facts and change strategies over dialogue stages; LLMs favor proposals and concessions, follow flatter trajectories, and display model-specific styles. Claude is descriptively closest to humans, Gemini is most skewed toward proposal, power, and residual behavior, and GPT-4o-mini uses more power and rights. Cooperative reciprocity is 88.5–98.2% for LLMs versus 73.7% for humans; GPT-4o-mini also reaches 48.8% competitive reciprocity versus 13.1%. These contrasts support the cautious conclusion that personality prompts do not make these systems reliable human behavioral proxies.

“Alignment,” however, is assessed mainly by comparing which coefficients cross p<.05 in separate regressions. Shared nonsignificance is not evidence of equivalence, and significance in one group but not another does not establish an effect difference. There is no pooled system-by-trait interaction, coefficient-difference interval, equivalence test, or multiplicity correction across hundreds of comparisons. Dyadic clustering is not documented even though each conversation contributes two related observations. Unequal sample sizes of 500/250/250/248 yield unequal power, and temperature 1 without seeds or repeated runs leaves simulation variance unknown.

IRP annotation uses GPT-4o-2024-08-06 at temperature 1. Three annotators, including an author, abandon direct classification after low agreement and instead judge GPT predictions as correct or incorrect on 25 human conversations. The final system reports 81% accuracy, .79 macro F1, and .81 weighted F1, with Positive Expectations at .69. This partial validation covers human text only and may not transfer equally to GPT-4o-mini, Claude, and Gemini styles; using an annotator from the same model family as one generator also permits style bias.

The official repository at commit 811840284ae06b689655bfa71a2da72bf00403e1 cannot reproduce the tables. It omits the 1,000 simulations, annotations, results, seeds, and raw human CSV. Its only processed CSV has 440 observations, 220 dyads rather than the reported 248, and 58 missing acceptance values. The required src/llm_api/base.py module, statsmodels dependency, and google-genai package used by the code are absent; there are no tests, CI, lockfile, or license despite an MIT claim. Several documented paths and pipeline steps do not exist, arguments and defaults conflict, and the appendix mistakenly describes the Gemini and GPT datasets as generated by two Claude models. The artifact exposes part of the design but is not an end-to-end executable or a verification package for the published results.

Español

Este trabajo pregunta si agentes LLM condicionados con rasgos Big Five reproducen las relaciones entre personalidad y conducta observadas en personas durante un conflicto. Compara una submuestra declarada de 248 diálogos humano-humano de KODIS con 1.000 negociaciones entre dos instancias del mismo modelo: 500 con GPT-4o-mini y 250 con Claude 3.7 Sonnet y Gemini 2.0 Flash. Todos negocian el mismo conflicto comprador-vendedor por una camiseta, con cinco decisiones sobre reembolso, reseñas y disculpas. Los resultados finales son puntuación, quién acepta y si se evita abandonar; las estrategias se clasifican con el marco Interests-Rights-Power en cooperación, neutralidad, competición y residual, además de reciprocidad, escalada y desescalada.

Los agentes reciben 15 adjetivos, tres por rasgo, elegidos de 70 pares bipolares y modificados para representar seis niveles. Sus distribuciones de rasgos imitan marginalmente la muestra humana, pero la medición no es equivalente: las personas contestan diez ítems BFI y los LLM reciben etiquetas discretas en un prompt. Además, el diseño introduce de antemano una asociación humana: el peso de la disculpa se hace depender de la agradabilidad usando una regresión de KODIS (B=2,13; p=.02), mientras los demás pesos se sortean. Por tanto, cualquier efecto de agradabilidad sobre resultados mezcla el supuesto mecanismo psicológico con una preferencia de utilidad programada por los autores.

En KODIS, la puntuación no presenta efectos de personalidad; para aceptar, mayor neuroticismo propio se asocia negativamente (B=-0,26; p=.026) y el de la pareja positivamente (B=0,27; p=.025); evitar abandonar tampoco presenta efectos. Los LLM muestran conjuntos distintos de coeficientes significativos según modelo y resultado. En estrategia, las personas usan sobre todo hechos y cambian de fase durante la conversación; los LLM recurren más a propuestas y concesiones, siguen trayectorias más planas y se comportan de modo específico al modelo. Claude es el más parecido descriptivamente al patrón humano, Gemini el más sesgado hacia propuesta, poder y residual, y GPT-4o-mini usa más poder y derechos. La reciprocidad cooperativa es 88,5–98,2 % en LLM frente a 73,7 % en personas; GPT-4o-mini también eleva la reciprocidad competitiva a 48,8 % frente a 13,1 %. Estos contrastes apoyan la conclusión cauta de que los prompts de personalidad no convierten a estos modelos en sustitutos conductuales fiables.

Sin embargo, “alineación” se decide principalmente comparando qué coeficientes cruzan p<.05 en regresiones separadas. Compartir ausencia de significación no demuestra equivalencia, y que un efecto sea significativo en un grupo y no en otro no demuestra que ambos efectos difieran. No se ajusta un modelo conjunto con interacción sistema×rasgo, no se aportan diferencias de coeficientes con intervalos ni pruebas de equivalencia, y no hay corrección por los cientos de contrastes. Tampoco se documenta agrupamiento por díada, pese a que cada conversación aporta dos participantes relacionados. Los tamaños 500/250/250/248 producen potencias distintas y la temperatura 1 sin semillas ni repeticiones no cuantifica variabilidad de simulación.

La anotación IRP usa GPT-4o-2024-08-06 a temperatura 1. Tres anotadores, uno autor, dejan de clasificar directamente tras obtener bajo acuerdo y pasan a juzgar las predicciones de GPT como correctas o incorrectas en 25 conversaciones humanas. El sistema alcanza 81 % de accuracy, F1 macro .79 y F1 ponderado .81; Positive Expectations queda en .69. Esa validación parcial solo cubre texto humano y puede no transferirse por igual a los estilos de GPT-4o-mini, Claude y Gemini; usar un anotador de la misma familia que uno de los generadores también admite sesgo de estilo.

La auditoría del repositorio oficial, commit 811840284ae06b689655bfa71a2da72bf00403e1, impide reproducir las tablas. No publica las 1.000 simulaciones, sus anotaciones, resultados, semillas ni el CSV humano bruto. El único CSV procesado contiene 440 observaciones, 220 díadas, no las 248 declaradas, y 58 valores ausentes para aceptación. Faltan el módulo src/llm_api/base.py, statsmodels en requirements y el paquete google-genai que exige el código; tampoco hay tests, CI, lockfile o licencia, aunque el README afirma MIT. Varias rutas y pasos descritos no existen, los argumentos y defaults discrepan, y el apéndice atribuye por error los datasets Gemini y GPT a dos Claude. El artefacto sirve para inspeccionar parte del diseño, no para ejecutar de extremo a extremo ni verificar los resultados publicados.

Research question

Do GPT-4o-mini, Claude 3.7 Sonnet, and Gemini 2.0 Flash, when given Big Five profiles by adjectives, reproduce the associations between own and partner personality, negotiation outcomes, and conflict strategies observed in comparable human-human dialogues?

Method

248 KODIS dialogues with personality for both parties are filtered and 1,000 LLM-LLM dialogues of the same scenario are simulated. LLM profiles sample six levels per trait and 15 adjectives; agreeableness determines the apology weight and the other weights are random. GPT-4o annotates IRP strategies. OLS or logistic regression relate ten personality predictors (five own and five partner) and role to three outcomes and multiple strategic metrics. The comparison between humans and models is interpreted from descriptive patterns and significance in separate models.

Sample: The paper starts from 4,061 KODIS participants and declares selecting 248 human-human dialogues with complete personality for both parties, removing missing values per analysis. It simulates 500 GPT-4o-mini dyads and 250 of Claude and Gemini, up to 25 rounds. IRP validation covers 25 human dialogues. The official release contains only one processed CSV with 440 rows, 220 buyers and 220 sellers; 58 rows lack do_accept_first, so it does not match the 248 declared dialogues nor allow reconstructing filters.

Findings

  • In KODIS, score and not-walk-away show no significant personality effects; accepting is associated with negative own neuroticism (B=-0.26; p=.026) and positive partner neuroticism (B=0.27; p=.025).
  • Significant personality coefficients in LLM outcomes differ between GPT-4o-mini, Claude, and Gemini and do not stably replicate the human pattern.
  • Humans use more facts and a more marked temporal progression; LLMs use more proposals and concessions and maintain flatter and model-specific patterns.
  • Claude is descriptively closer to human strategies; Gemini presents more proposal, power, and residual and does not use Positive Expectations or Procedural; GPT-4o-mini maintains more power and rights.
  • Cooperative reciprocity is higher in LLMs (88.5-98.2 %) than in KODIS (73.7 %); competitive reciprocity reaches 48.8 % in GPT-4o-mini versus 13.1 % in humans.
  • IRP annotation validated on 25 human dialogues obtains 81 % accuracy, macro F1 .79 and weighted .81; Positive Expectations has F1 .69.
  • The evidence supports caution regarding the use of personality agents as human proxies, but does not directly quantify a distance or equivalence of effects between systems.

Limitations

  • A single e-commerce role-play does not represent real legal mediation, clinical, labor, family, or intercultural conflict, despite the language of high-impact applications.
  • Humans respond to BFI-10 and models receive discretized adjectives; there is no measurement invariance or equivalence between the observed human trait and the prompt manipulation.
  • There is no manipulation check in these dialogues confirming that each agent expresses the assigned profile or that the five traits remain distinguishable during negotiation.
  • LLM profiles reproduce marginal distributions, not one-to-one pairings of persons, couples, priorities, and conversations. "Matched" means scenario and distribution, not individual replicas.
  • The design makes the importance of the apology depend on agreeableness using a regression from the same human source. This structurally introduces a trait-utility pathway before observing behavior.
  • The other priorities are drawn at random, so human and LLM payoffs do not come from the same mechanism and the score comparison is not a clean validation of personality.
  • The prompts explicitly prescribe starting with the most important item, conceding, reciprocating, switching after a stalemate, negotiating at least three issues, and avoiding no agreement; these rules produce part of the evaluated strategies.
  • Few-shot examples and temporal warnings favor proposals, concessions, and agreements, making it difficult to attribute those behaviors to the model or to personality.
  • Results are fitted separately. A coefficient being significant in one set and not in another does not prove a difference; shared non-significance does not prove alignment either.
  • There is no joint model with a system x trait interaction, coefficient contrast, difference interval, equivalence test, or aggregate alignment metric.
  • Hundreds of tests are run at 5 % without multiplicity correction, which increases false positives. P-values printed as .000 should be reported as <.001.
  • The sizes 500, 250, 250, and 248 imply different power. The claim that each count gives statistically valid estimates is not supported by power or precision analyses.
  • Each dyad contributes two participants and related turns. HC3 corrects heteroscedasticity in part of the OLS, but does not document errors clustered by conversation or multilevel modeling.
  • Logistic regressions of outcomes do not use robust errors; log-odds coefficients are presented as B without odds ratios or marginal effects, making it difficult to compare magnitudes.
  • Ten correlated personality predictors are included at once without multicollinearity diagnostics, regularization, or coefficient stability checks.
  • Analyses remove missing values per variable, with changing N and no exclusion flow, missingness pattern, or sensitivity analysis.
  • Temperature 1, absence of seeds, and a single realization per profile prevent separating sampling variance of the model from trait effects.
  • No dated snapshots of the three generators or complete logs are published. The article abbreviates GPT-4o-mini as GPT-4, creating identity ambiguity.
  • Only same-model dyads are studied. There are no pairs between models, human-LLM, conditions without personality, contradictory profiles, or comparison with simple style instructions.
  • Human annotation drops direct classification after low agreement and moves to validating GPT predictions, which may anchor judges and does not estimate independent agreement on IRP labels.
  • Only 25 human conversations are validated, not LLM outputs. GPT-4o performance as annotator may vary by generator model and favor the style of its own family.
  • An accuracy of 81 % and F1 .69 in Positive Expectations introduce measurement error not propagated to the strategic intervals or regressions.
  • Claims of flexibility, rigidity, adaptation, or context are inferred from aggregate frequencies; they are not tested as independent psychological constructs.
  • Claude's greater descriptive similarity does not equate to human validity: it may reflect a more factual textual style induced by model or prompt.
  • Robustness to order, wording, intensity, adjective selection, examples, or rules is not evaluated; the paper itself acknowledges this limitation.
  • The official repository does not contain simulations, complete annotations, results, or the raw KODIS subset; it does not allow regenerating any main table.
  • The released CSV has 220 dyads, 28 fewer than the paper, and 58 observations without acceptance. There are no IDs or provenance to explain the difference.
  • src/llm_api/base.py is missing, so the simulation pipeline does not import; requirements omit statsmodels and declare google-generativeai while the code uses google.genai from google-genai.
  • The README refers to nonexistent files, scripts, and outputs; n_exp and n_round defaults disagree and type=bool prevents disabling personality correctly.
  • The personality_setting=False condition fails because the code keeps using player1_selected_traits and player2_selected_traits that are not defined.
  • prompt_build_v2 evaluates template contents as Python, an unsafe and unnecessary practice that also fragilizes reproduction.
  • There is no lockfile, fixed environment, tests, CI, or license in the repository; claiming MIT in README does not by itself grant a verifiable license.
  • The appendix contains copy errors: it describes the Gemini and GPT-4o-mini datasets as generated by two Claude Sonnet 3.7.
  • Consent, ethical approval, or reuse conditions for this subsample are not documented; the paper relies on the governance of the prior KODIS corpus.

What the study does not establish

  • It does not demonstrate that GPT-4o-mini, Claude, or Gemini embody a stable human personality; it tests responses to adjectival profiles in a single script.
  • It does not establish statistical equivalence or divergence of effects between humans and LLMs through direct contrasts; it compares significance in separate models.
  • It does not demonstrate that shared absence of p<.05 is behavioral alignment or that exclusive significance of one group is a difference between groups.
  • It does not causally identify that personality produces the score differences, because the utility function explicitly depends on agreeableness.
  • It does not validate the use of these agents in legal mediation, real negotiation, or socially sensitive decisions.
  • It does not demonstrate that Claude is a reliable human proxy; it is only more similar in some distributions and descriptive trajectories.
  • It does not prove generalization to other scenarios, languages, cultures, models, prompts, configurations, or human-AI interactions.
  • It does not allow reproducing the simulations, annotations, or tables from the available official repository.

Traceability

Scope: Full text

Version: arXiv:2602.07414v1, submitted 7 February 2026; AAAI 2026 AISI Special Track, 19 pages

Consulted source: https://arxiv.org/pdf/2602.07414v1

Review: Codex full-text, bilingual-fidelity, 19-page visual, arXiv-v1, AAAI-status, construct-validity, BFI-measurement-equivalence, prompt-intervention, circular-utility, dyadic-dependence, multiple-testing, coefficient-comparison, annotation-transfer, data-release, code-executability, dependency, documentation and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o-mini, 500 same-model L2L dialogues
  • Claude 3.7 Sonnet, 250 same-model L2L dialogues
  • Gemini 2.0 Flash, 250 same-model L2L dialogues
  • GPT-4o-2024-08-06 IRP annotator, run 19 March 2025

Instruments and metrics

  • Ten-item Big Five Inventory data in KODIS
  • Six-point polarity-degree LLM personality profile
  • Seventy bipolar adjective pairs, 15 adjectives per agent
  • Interests-Rights-Power strategy framework
  • Outcome score, accept and not-walk-away
  • IRP ratio, reciprocity, escalation and de-escalation
  • OLS and logistic regression with self traits, partner traits and role
  • A-Kappa, accuracy, macro F1 and weighted F1 for annotation

Data used

  • KODIS human-human subset, reported 248 dyads
  • Official processed KODIS release, 220 dyads and 440 participant rows
  • GPT-4o-mini L2L simulations, 500 reported and unreleased
  • Claude 3.7 Sonnet L2L simulations, 250 reported and unreleased
  • Gemini 2.0 Flash L2L simulations, 250 reported and unreleased
  • Twenty-five-conversation human IRP evaluation subset

Evidence and location

  • Metadata, abstract, question, and contributions: arXiv:2602.07414v1, pp. 1-2, Abstract and Introduction
  • KODIS, 248 dialogues, and measures: Paper, pp. 3-4, KODIS and Behavioral Measures
  • Profiles, 70 pairs, six levels, and simulations: Paper, pp. 4-5, L2L Dataset Construction and LLM Personality Profile
  • Agreeableness programmed into apology importance: Paper, p. 5, LLM Personality Profile; B=2.13, p=.02
  • Statistical models and outcome effects: Paper, pp. 5-6 and supplementary pp. 11-19, regression tables
  • IRP distribution, reciprocity, and temporal dynamics: Paper, pp. 6-8, Figures 2-4 and Results and Discussion
  • Declared limitations and conclusion: Paper, p. 8, Conclusion and Limitations and Future Work
  • IRP annotation validation: Paper supplementary, pp. 9-11, Tables 8-9; 25 conversations, accuracy .81, macro F1 .79
  • Hyperparameters, matching, and dataset errors: Paper supplementary, pp. 12-13, Hyperparameters, Personality Matching and Dataset Statements
  • Integral visual inspection: Paper, all 19 rendered pages, including every figure, table, prompt, and appendix page
  • Code, released data, and reproducibility: Official repository commit 811840284ae06b689655bfa71a2da72bf00403e1, all 36 tracked files
  • 220 released dyads and 58 missing acceptances: Official repository, data/KODIS/KODIS_H2H_processed.csv; 440 rows split equally by role
  • Incomplete pipeline and inconsistent dependencies: Official repository, requirements.txt, src/llm_api, scripts, and README.md at audited commit