AI-exhibited Personality Traits Can Shape Human Self-concept through Conversations

Applications, bias, and safety2026arXivApproved editorial review

Authors: Jingshu Li, Tianqi Song, Nattapat Boonprakong, Zicheng Zhu, Yitian Yang, Yi-Chieh Lee

Keywords: Large Language Models, Personality, Bias, Persona, AI Safety

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

6
Authors
8
Findings
35
Limitations
14
Evidence

Editorial summary

English

This experiment asks whether a person's reported self-concept moves closer, after a conversation, to a 20-adjective vector assigned to GPT-4o. It recruits 110 US adults through CloudResearch Connect; 92 complete the study, leaving 46 per condition. Before and after chatting, they rate themselves from 0 to 100 on twelve positive adjectives, such as cheerful, enthusiastic, confident, and sincere, and eight negative ones, selfish, cranky, jealous, whiny, stingy, and others. Participants are randomized to a personal or non-personal topic list and converse for 5–15 minutes and at least ten turns with the same GPT-4o and system prompt. The exact model snapshot, API date, and effective parameters are not reported.

“AI personality” is neither manipulated nor established through independent behavior: the same GPT-4o is asked to complete the adjective scale 100 times, and item averages define its vector. The resulting profile is strongly socially desirable: high on trustworthy (.873), sincere (.850), and loyal (.814), and low on all eight negative attributes (.090–.171). The primary outcome is the reduction in Manhattan distance between each human's 20 self-ratings and this vector. Across the full sample, the mean reduction is .251 aggregate units (d=.509, p<.001). The most informative randomized contrast is the time-by-topic interaction: F(1,89)=5.109, p=.026, partial eta squared=.054. In the personal condition, total distance falls from 4.188 to 3.832 (d=-.216), or 1.78 scale points per adjective; in the non-personal condition it falls from 4.545 to 4.384 and does not survive the Bonferroni adjustment (p=.279).

This supports a small shift toward the selected vector after personal rather than non-personal conversation. It does not identify GPT-4o personality as the cause: topic, not chatbot personality, is randomized. Personal prompts elicit autobiography, gratitude, pride, friendship, and emotional reflection, whereas non-personal prompts concern technology, animals, and speculative scenarios. The effect can therefore reflect self-affirmation, social desirability, repeated measurement, regression to the mean, or positive conversational content. Attributing it to AI traits would require contrasting chatbot personalities, a trait-neutral chatbot, and ideally a no-conversation control.

The duration result is correlational: r=.245, p=.019. Participants choose whether to continue after five minutes, so engagement, enjoyment, or susceptibility may cause both longer conversations and more change. Of 46 people in the personal condition, 33-72%, report noticing no self-concept change. The homogenization claim compares all 1,035 pre/post pairwise distances among those same 46 people: 5.319 versus 5.065. Its t(1034) and Wilcoxon tests treat participant-overlapping pairs as independent observations; this is pseudoreplication and makes p<.001 far more precise than an effective sample of 46 supports. The study also does not test whether homogenization changes more than in the non-personal arm.

With N=92, the mediation model associates alignment with total enjoyment (beta=.951, p=.008) and proposes alignment→perception accuracy→shared reality→enjoyment (indirect beta=.259, p=.014; direct p=.651). Mediators and outcome are all collected in the same post-survey without experimental temporal ordering. Alignment and “accuracy” share the same AI reference vector and item scale; participants also rate their post-conversation self first and the AI later, which the authors acknowledge can anchor responses. The SEM describes covariance, not causal mediation or proof that enjoyment follows self-concept change.

The instrument is not a Big Five scale or a comprehensive personality measure; it selects adjectives expected to change in brief interactions. Comparing within-person change to between-person differences does not establish construct validity, and the r=.954 correlation between human and AI means across 20 items can be driven by shared valence. The descriptive table shows a broad positive self-rating shift, several positive attributes rise and nearly all negative ones fall, which is compatible with self-affirmation rather than imitation of a distinct personality.

The paper is useful as a warning and an initial experimental signal that conversational context can move momentary self-ratings. Its authors describe the effect as state-like and acknowledge that persistence is unknown. No public data, anonymized chats, analysis code, preregistration, randomization details, seed, or executable artifact is linked from the paper or arXiv, so filters, scoring, SEM, and sensitivity analyses cannot be independently audited. Proposals about mass manipulation, mental health, minors, cultural erosion, or improved well-being are design scenarios rather than findings from this short experiment.

Español

Este experimento estudia si el autoconcepto declarado por una persona se acerca, tras una conversación, a un vector de 20 adjetivos atribuido a GPT-4o. Recluta 110 adultos estadounidenses en CloudResearch Connect; 92 completan el estudio y quedan 46 por condición. Antes y después de conversar, puntúan de 0 a 100 doce adjetivos positivos, por ejemplo alegre, entusiasta, confiado o sincero, y ocho negativos, egoísta, malhumorado, celoso, quejica, tacaño, etc.. Se les asigna al azar una lista de temas personales o no personales y conversan entre 5 y 15 minutos, con al menos diez turnos, con el mismo GPT-4o y el mismo prompt. El snapshot del modelo, la fecha de API y los parámetros efectivos no se especifican.

La “personalidad de la IA” no se manipula ni se observa mediante conducta independiente: se estima pidiendo al propio GPT-4o que complete 100 veces la misma escala y promediando cada adjetivo. El vector resultante es muy socialmente deseable: alto en trustworthy (.873), sincere (.850) y loyal (.814), y bajo en los ocho atributos negativos (.090–.171). La variable principal es la reducción de distancia Manhattan entre las 20 autoevaluaciones humanas y ese vector. En toda la muestra, la reducción media es .251 unidades agregadas (d=.509; p<.001). El contraste aleatorizado más informativo es la interacción tiempo×tema: F(1,89)=5,109, p=.026, η²p=.054. En temas personales la distancia total baja de 4,188 a 3,832 (d=-.216), equivalente a 1,78 puntos de escala por adjetivo; en temas no personales baja de 4,545 a 4,384 y no supera el ajuste Bonferroni (p=.279).

Esto respalda que una conversación personal produce un pequeño desplazamiento de respuestas hacia el vector elegido, comparada con una conversación no personal. No identifica que la personalidad de GPT-4o sea la causa: la aleatorización cambia el tema, no el chatbot. Las listas personales inducen autobiografía, gratitud, orgullo, amistad y reflexión emocional, mientras las no personales tratan tecnología, animales o escenarios especulativos. El resultado puede reflejar autoafirmación, deseabilidad social, repetición de la escala, regresión a la media o el contenido positivo de la conversación. Para atribuirlo a rasgos de IA harían falta al menos personalidades contrastadas, un chatbot sin señales de rasgo y, preferiblemente, un control sin conversación.

La asociación con duración es correlacional: r=.245, p=.019. Después de cinco minutos cada persona decide si continúa, por lo que interés, disfrute o susceptibilidad pueden causar tanto más turnos como mayor cambio. De las 46 personas en temas personales, 33, 72 %, dicen no haber notado cambios. La afirmación de homogeneización compara las distancias pre/post de los 1.035 pares posibles entre esas mismas 46 personas: 5,319 frente a 5,065. El t(1034) y Wilcoxon tratan pares que comparten participantes como observaciones independientes; es pseudorreplicación y hace que p<.001 parezca mucho más preciso que el tamaño efectivo de 46 personas. Tampoco se contrasta si la homogeneización cambia más que en la condición no personal.

El modelo de mediación con N=92 asocia alineación con disfrute total (β=.951; p=.008) y propone la cadena alineación→precisión percibida→realidad compartida→disfrute (efecto indirecto β=.259; p=.014; efecto directo p=.651). Todos los mediadores y el resultado se recogen en la misma encuesta posterior, sin orden temporal experimental. Alineación y “precisión” comparten el mismo vector de IA y la misma escala; además, los participantes valoran primero su autoconcepto post y después a la IA, lo que los autores reconocen como posible anclaje. El SEM describe covariación, no prueba mediación causal ni que el disfrute sea consecuencia de un cambio de autoconcepto.

La escala no es Big Five ni una medición completa de personalidad: selecciona adjetivos especialmente susceptibles a cambio breve. Comparar cambio intraindividual con diferencias entre personas no establece validez de constructo; la correlación r=.954 entre medias de 20 ítems humanos y de IA puede estar dominada por su valencia común. La tabla muestra un patrón general de autoevaluación más positiva, suben varios atributos positivos y bajan casi todos los negativos, que es compatible con autoafirmación y no exclusivo de imitación de una personalidad.

El artículo es valioso como alerta y como primera señal experimental de que el contexto conversacional puede mover autoevaluaciones momentáneas. Sus propios autores califican el efecto como state-like y reconocen que no saben si persiste. No hay datos, chats anonimizados, código de análisis, preregistro, semilla, especificación de aleatorización ni artefacto ejecutable enlazado desde el paper o arXiv; por ello no se pueden auditar filtros, escalas, SEM o sensibilidad. Las propuestas sobre manipulación masiva, salud mental, menores, erosión cultural o mejora del bienestar son escenarios de diseño, no resultados de este experimento breve.

Research question

Do self-evaluations of 20 traits change after conversing with GPT-4o toward the adjective vector that the model itself declares, does that approach change more with personal topics, and is it related to homogeneity, duration, and enjoyment?

Method

Mixed pre/post online experiment with 92 adults from the United States randomly assigned to personal or non-personal topics. All converse 5–15 minutes and at least ten turns with the same default GPT-4o. Humans and model respond to the same scale of 20 adjectives. Alignment is the reduction of Manhattan distance to the average of 100 responses per item from the model. t/Wilcoxon, repeated-measures ANOVA with turns as covariate, correlation, comparison of distances between all pairs, and a serial mediation SEM are used.

Sample: 110 people were recruited; 18 did not complete and 92 were analyzed, 46 per condition. They resided in the United States, were 21–60 years old, mean age 39.0 (SD 10.1), 43 women and 49 men; 63% had at least an associate degree. The planned session lasted 15 minutes and was compensated with 2.50 USD. The conversation lasted 5 to 15 minutes, mean 9.0 (SD 3.9), with at least ten turns and mean 20.9 (SD 9.0). Attrition per arm, reasons for dropout, complete baseline balance, and randomization method are not reported.

Findings

  • Across the entire sample, the mean reduction in distance to the GPT-4o vector is .251 (SD .493), t(91)=4.481, p<.001, d=.509; the analysis mixes both topics.
  • The time×topic interaction is F(1,89)=5.109, p=.026, η²p=.054. The personal change is small, d=-.216; the non-personal is not significant after Bonferroni, p=.279.
  • In personal topics the aggregated distance drops from 4.188 to 3.832, a reduction of .356 over 20 adjectives; in non-personal topics it drops .161.
  • More turns correlate with greater alignment, r=.245, p=.019, but duration is chosen by participants after the minimum and is not randomized.
  • Only 13 of 46 personal participants report noticing a change; 33, 72%, do not perceive it.
  • Pairwise distances in the personal arm drop from 5.319 to 5.065, but the inference uses 1.035 non-independent pairs derived from 46 people.
  • The SEM reports a total effect on enjoyment β=.951, p=.008 and a serial indirect effect β=.259, p=.014; the subsequent cross-sectional data do not establish causal mediation.
  • The profile that GPT-4o assigns to itself is strongly positive and human self-reports generally shift toward more socially desirable attributes.

Limitations

  • The topic is randomized, not the AI personality. All interact with the same chatbot, so there is no causal contrast of traits.
  • A no-conversation condition, another GPT-4o personality, and a neutral chatbot are missing; personality, positivity, conversation, and scale repetition are not separated.
  • Personal and non-personal topics differ in self-reference, emotion, valence, depth, and demand for self-affirmation, not only in exposure to traits.
  • The AI vector comes from the model's own self-report in a separate task, not from independent coding of the responses each participant received.
  • It is not demonstrated that the 100 scale responses correspond to the traits effectively expressed in each individual conversation.
  • The model is only identified as GPT-4o with defaults; snapshot, API date, effective temperature, top-p, seed, and backend variability are missing.
  • The list of 20 adjectives is not Big Five, has no reported factor structure, and mixes affective states, social virtues, and negative traits.
  • Selecting items by their susceptibility to brief change favors finding pre/post change and limits interpreting the result as stable personality.
  • The AI profile is markedly socially desirable; approaching it may be a generic improvement in self-evaluation, not imitation of a specific personality.
  • The Manhattan distance weights all adjectives equally and collapses direction and content into a single value, hiding which dimensions generate the effect.
  • The personal reduction of .356 aggregated units equals 1.78 points per item on a 0–100 scale; the title does not communicate this small magnitude.
  • H1 tests alignment on the 92 combined people although the interpreted effect is localized in the personal arm.
  • There is no non-AI control or test-retest to estimate regression to the mean, habituation, scale consistency, or social desirability.
  • The correlation with turns is exploratory and post-treatment. After five minutes continuing is a choice, so it does not prove that longer duration causes alignment.
  • The 1.035 distances of H3 repeatedly share the same 46 people; t and Wilcoxon violate independence and produce pseudoreplication.
  • H3 does not directly test the change in homogeneity between conditions; it analyzes only the selected group because it previously showed significance.
  • The decrease in distance between pairs is .254 aggregated, about 1.27 points per adjective, and does not demonstrate durable loss of social or cultural diversity.
  • The correlation of 4.950 pairs of GPT measurements also uses dependent comparisons; the significance of each correlation does not prove trait validity.
  • The correlation r=.954 uses only 20 aggregated means and may reflect shared valence or stereotypes, not individual precision or construct validity.
  • Comparing intraindividual change with interindividual distance does not validate a scale; it only demonstrates that people differ more from each other than their own pre/post.
  • The SEM with N=92 is constructed after preliminary regressions and without an identified preregistration. Good fit does not resolve confounding, causal direction, or overfitting.
  • Alignment and perceived accuracy share the same AI vector and the same 20 items, introducing mathematical coupling into the mediation.
  • Mediators and enjoyment are measured in a single post-survey; there is no manipulation or temporal sequence that identifies the causal chain.
  • Inferring total mediation because the indirect effect is significant and the direct effect is not does not prove the absence of alternative pathways.
  • The fixed order post-self-concept before perceived AI traits may anchor both scales; the authors acknowledge it affects H5.
  • Of 110 recruited, 18 do not complete. Attrition is not broken down by condition, reason, moment, or differences, preventing assessment of dropout bias.
  • The manipulation is not perceived according to assignment by 14 of 92 people; the check is reported but not per-protocol sensitivity.
  • Power is generically based on previous medium-large effects and N=90, without a target effect, primary hypothesis, or adjustment for attrition and multiplicity.
  • No global correction for five hypotheses and additional analyses is reported; Bonferroni is limited to post-hoc comparisons of H2.
  • The textual count is inconsistent: it is stated that 13 turns mention traits, but the breakdown 5 non-personal + 7 personal sums to 12.
  • The manual analysis of 1.919 turns does not publish protocol, labels, agreement between the two researchers, or anonymized logs.
  • A single brief exchange with U.S. adults does not allow inferring persistence, accumulation, minors, disorders, other cultures, or everyday use.
  • The paper mentions IRB approval and consent, but gives no identifier, risk protocol, or reproducible treatment of sensitive conversational data.
  • No code, data, preregistration, or executable official material linked was found; randomization, exclusions, distance calculations, or SEM cannot be verified.
  • The recommendations to limit duration, profile vulnerability, or design positive traits are normative proposals and require their own ethical evaluation.

What the study does not establish

  • It does not establish that the personality of GPT-4o causes the change: that personality is not manipulated and only the topic is randomized.
  • It does not demonstrate a real or persistent personality change; it measures a small immediate shift in self-evaluations of adjectives.
  • It does not demonstrate that the chatbot exhibited during each conversation the same vector it declared in separate questionnaire responses.
  • It does not exclude self-affirmation, social desirability, scale repetition, regression to the mean, thematic valence, or expectations about AI.
  • It does not causally prove that longer conversations produce more alignment or that alignment produces enjoyment.
  • It does not prove the causal chain of perception, shared reality, and enjoyment; the SEM uses correlational post measures.
  • It does not establish population homogenization, cultural loss, or collective harm; H3 contains pseudoreplication and a small absolute change.
  • It does not validate interventions for mental health, education, companionship, minors, or vulnerable people.
  • It does not allow reproducing the results from identified public artifacts.

Traceability

Scope: Full text

Version: arXiv:2601.12727v1, submitted 19 January 2026; published at ACM CHI 2026, DOI 10.1145/3772318.3790654, 20 pages

Consulted source: https://arxiv.org/pdf/2601.12727v1

Review: Codex full-text, bilingual-fidelity, 20-page visual, arXiv-v1, ACM-DOI, randomized-design, causal-identification, self-concept-construct, adjective-valence, measurement-validity, Manhattan-distance, effect-size, pairwise-dependence, pseudoreplication, SEM-mediation, attrition, model-snapshot, privacy and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o, unspecified snapshot, participant-facing chatbot
  • GPT-4o, same unspecified snapshot, 100 repeated adjective-scale completions per item

Instruments and metrics

  • Twenty-adjective self-concept slider scale, 0–100
  • Twelve positively valenced and eight negatively valenced traits
  • Manhattan human-AI self-concept distance
  • Manhattan inter-participant self-concept distance
  • Single-item conversation enjoyment, 1–7
  • Eight-item adapted Generalized Shared Reality scale
  • Perceived AI adjective ratings
  • Repeated-measures ANOVA and Bonferroni pairwise comparisons
  • Structural equation serial mediation model

Data used

  • CloudResearch Connect randomized experiment, 92 completed participants
  • Personal-topic arm, 46 participants and 1,035 dependent participant pairs
  • Non-personal-topic arm, 46 participants
  • 1,919 unreleased human-AI conversation turns
  • Two thousand GPT-4o scale responses summarized as 100 repetitions per trait

Evidence and location

  • Metadata, abstract, version, and DOI: arXiv:2601.12727v1 abstract page and ACM DOI 10.1145/3772318.3790654
  • Hypotheses and declared causal scope: Paper, pp. 5–6, Hypotheses Development
  • Sample, power, randomization, and compensation: Paper, pp. 5–6, Participants and Procedure and Conditions
  • Chatbot, topics, and parameters: Paper, p. 6 and Appendix A–B, Experimental Interface, topic lists and complete prompts
  • Scale, distances, and AI vector: Paper, pp. 7–9, Measurements and Derived Measurements; Appendix C–D
  • Manipulation check and scale validation: Paper, p. 9, Sections 5.1–5.2
  • Alignment and topic effect: Paper, pp. 9–11, Sections 5.3–5.3.1, Table 2 and Figure 3
  • Homogenization and pairwise dependence: Paper, pp. 11–12, Section 5.4 and Figure 4; 1,035 pairs from 46 participants
  • SEM and mediation: Paper, pp. 12–13, Section 5.5, Table 3 and Figure 5
  • Declared limitations and design scenarios: Paper, pp. 13–16, Discussion, Design Implications and Limitations
  • Descriptives per adjective: Paper, p. 20, Appendix D, Table 4
  • Sensitivity excluding personality question: Paper, pp. 15 and 20, Limitations and Appendix E, N=89
  • Comprehensive visual inspection: Paper, all 20 rendered pages, including every figure, table and appendix page
  • Absence of linked public artifacts: Paper and official arXiv article surfaces checked 15 July 2026; no data/code/preregistration link identified