The Personality Dimensions GPT-3 Expresses During Human-Chatbot Interactions

Personas, identity, and agents2024ACMApproved editorial review

Authors: Nikola Kovačević, Christian Holz, Markus Gross, Rafael Wampfler

Keywords: text-davinci-002, Human-Chatbot Interaction, Perceived Personality, Exploratory Factor Analysis, Conversational Agents, Big Five, Psychometrics, Interrater Reliability, Descriptor Elicitation, Reproducibility

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

4
Authors
14
Findings
24
Limitations
12
Evidence

Editorial summary

English

This study does not measure an intrinsic GPT-3 personality. It investigates how mostly university participants perceived personality in English conversations generated by text-davinci-002 in October-November 2022. In the first experiment, 86 participants interacted for three weeks with three personas and supplied 9,267 three-adjective self-reports. The system also injected one of six emotional states, one of four conversation starts and the user's own messages; the observed object is therefore a prompt-conditioned, co-produced conversation rather than the model in isolation.

The 27,801 adjective entries yielded 2,999 unique terms and then 147 descriptors through consequential construction choices: manual phrase reduction, manual translation of 104 terms, Merriam-Webster correction, removal of negation prefixes, screening against prior lists, greedy frequency-ordered synonym clustering and an elbow-chosen t=10 cutoff. The result is a curated instrument, not a neutral extraction or automated summary of participant language. The official supplement publishes all descriptors, loadings and Big Five correspondences, but not the intermediate decisions.

A second open survey recruited 556 people and retained 425 after excluding 131 (23.5%): nine for not using all four categories and 122 for finishing outside 8-25 minutes. Each participant read a different random set of three conversations and rated all 147 adjectives once for the bundle's overall chatbot personality. Exploratory factor analysis reports KMO=.896, but the retention rules do not select eight factors: Kaiser suggests 34 and the scree plot four or six. The authors analyze five for Big Five comparison, ten for earlier-work comparison and choose eight as an intermediate, interpretable value. That solution explains 38.2% of variance and is not confirmed in another sample.

The factors are subjectively labeled decency, profoundness, instability, vibrancy, engagement, neuroticism, serviceability and subservience. They form a plausible exploratory taxonomy of perceptions, but each vector mixes one rater and three different conversations; covariance can incorporate content, prompts, users, bundle selection and response styles. There is no multilevel analysis, parallel analysis, confirmation split, CFA, stability bootstrap or external replication.

The most prominent reliability figure requires an important correction. The main text presents alpha=.78 for people who saw the same three conversations, but Appendix B shows .57 without an ordering restriction and only 23 rater pairs. The .78 value additionally requires the same conversation to appear last and is based on nine pairs. Alpha=.31 among more than 80,000 pairs sharing no conversation is not interrater reliability of the same object; it is profile similarity across different stimuli and can reflect adjective base rates or response styles.

Big Five agreement is not convergent validation either. After list matching, a psychologist manually assigned 95 of 147 adjectives and 24 remained unmatched; “agreement” is the percentage of a factor's words coded to each trait, not correlation with a Big Five instrument administered on the same stimuli. Comparison with voice assistants likewise uses another study, sample and vocabulary; measurable overlaps are 27-41% and do not isolate technological progress.

The faithful conclusion is that participants in 2022 organized perceptions of these particular conversations into several interpretable social, emotional and functional dimensions. The study contributes an unusual longitudinal design, identifies the model and settings and publishes complete tables. It does not validate a stable GPT-3 personality, a universal eight-factor taxonomy, psychometric failure of Big Five, prompt independence or generalization to ChatGPT, GPT-4 or current models. Numerical reproduction is also impossible: official pages and the ETH record expose PDFs, BibTeX and metadata, but no conversations, rating matrix, exclusions, code, environment or immutable release.

Español

Este estudio no mide una personalidad intrínseca de GPT-3. Investiga cómo personas, en su mayoría universitarias, percibieron la personalidad mostrada en conversaciones inglesas generadas por text-davinci-002 en octubre-noviembre de 2022. En el primer experimento, 86 participantes conversaron durante tres semanas con tres personajes y aportaron 9.267 autoinformes de tres adjetivos. El sistema inyectaba además uno de seis estados emocionales, cuatro tipos de inicio y los mensajes del usuario; el objeto observado es, por tanto, una conversación coproducida y condicionada por prompts, no el modelo aislado.

De 27.801 entradas se obtuvieron 2.999 términos únicos y después 147 descriptores mediante una cadena con decisiones importantes: reducción manual de frases, traducción manual de 104 términos, corrección con Merriam-Webster, eliminación de prefijos de negación, cribado con listas previas, clustering codicioso de sinónimos ordenado por frecuencia y un umbral t=10 elegido por codo. El resultado es un instrumento construido y curado, no una extracción neutral ni un resumen automático del lenguaje de los participantes. El suplemento oficial publica los 147 descriptores, sus cargas y su correspondencia Big Five, pero no las decisiones intermedias.

En un segundo estudio abierto se reclutaron 556 personas y se conservaron 425 tras excluir 131 (23,5 %): nueve por no usar las cuatro categorías y 122 por completar fuera de 8-25 minutos. Cada participante leyó un conjunto aleatorio distinto de tres conversaciones y puntuó una sola vez los 147 adjetivos para la personalidad global de ese conjunto. El análisis factorial exploratorio tiene KMO=0,896, pero las reglas de retención no señalan ocho factores: Kaiser sugiere 34 y el scree cuatro o seis. Los autores analizan cinco para compararlo con Big Five, diez para compararlo con un trabajo anterior y eligen ocho como valor intermedio e interpretable. Esa solución explica el 38,2 % de la varianza y no se confirma en otra muestra.

Los ocho factores se etiquetan subjetivamente como decencia, profundidad, inestabilidad, vitalidad, implicación, neuroticismo, utilidad y servilismo. Forman una taxonomía exploratoria plausible de percepciones, pero cada vector mezcla un rater y tres conversaciones distintas; la estructura puede incorporar contenido, prompts, usuarios, selección del conjunto y estilos de respuesta. No hay análisis multinivel, parallel analysis, split de confirmación, CFA, bootstrap de estabilidad ni replicación externa.

La cifra de fiabilidad más llamativa necesita una corrección esencial. El texto principal presenta alpha=0,78 para quienes vieron las mismas tres conversaciones, pero el apéndice muestra que sin restricción de orden es 0,57 en solo 23 pares. El 0,78 aparece únicamente al exigir además que la última conversación sea la misma y queda en nueve pares. El alpha=0,31 entre más de 80.000 pares sin ninguna conversación compartida no es fiabilidad interjueces del mismo objeto: es semejanza de perfiles ante estímulos diferentes y puede reflejar frecuencias generales de los adjetivos o estilos de respuesta.

La relación con Big Five tampoco es validación convergente. Tras comparar con listas previas, un psicólogo asignó manualmente 95 de los 147 adjetivos y 24 quedaron sin correspondencia; el “agreement” es el porcentaje de palabras de cada factor codificadas en cada rasgo, no una correlación con un instrumento Big Five administrado sobre los mismos estímulos. Del mismo modo, la comparación con asistentes de voz usa otro estudio, otras muestras y otros vocabularios; los solapamientos mensurables son 27-41 % y no aíslan un avance tecnológico.

La conclusión fiel es que participantes de 2022 organizaron sus percepciones de estas conversaciones concretas en varias dimensiones sociales, emocionales y funcionales interpretables. El trabajo aporta un diseño longitudinal poco común, identifica el modelo y los parámetros y publica las tablas completas. Pero no valida una personalidad estable de GPT-3, una taxonomía universal de ocho factores, el fracaso psicométrico de Big Five, independencia respecto del prompt o generalización a ChatGPT, GPT-4 o modelos actuales. Tampoco puede reproducirse numéricamente: las páginas oficiales y el registro ETH ofrecen PDFs, BibTeX y metadatos, pero no conversaciones, matriz de ratings, exclusiones, código, entorno o release inmutable.

Research question

What vocabulary do participants use to describe the perceived personality of a chatbot based on text-davinci-002 during prolonged interaction, what exploratory factorial structure appears when scoring that vocabulary over conversations, and how does it overlap lexically with Big Five and a previous voice assistant model?

Method

Design in two stages. First, 86 participants interacted for three weeks with text-davinci-002 using three personas, six random emotions and four conversation starts; they produced 1,998 conversations and 9,267 reports of three adjectives. A manual, lexical and deterministic chain reduced 27,801 entries to 147 descriptors. Afterwards, 556 people took an open survey; 425 remained after exclusions and each one scored the 147 descriptors once after reading three random conversations. EFA was applied with oblimin and solutions of 5, 8 and 10 factors, ten Berge scores, pairwise Krippendorff alpha and lexical overlap with Big Five and Völkel et al.

Sample: Interaction experiment: 86 people, 42 women and 44 men, 18-41 years (mean 25.4; SD 3.9), mostly students from ETH/University of Zurich; 87% with English C1+ and 58% with chatbot experience. Survey: 556 recruited, 324 men, 230 women and 2 other, 17-50 years (mean 22.9; SD 3.6), mostly ETH students; 89% C1+ and 59% with experience. 131 were excluded and 425 remained.

Findings

  • 1,998 conversations, 9,267 self-reports and 27,801 adjective entries were collected; 2,999 unique terms were reduced to 147 descriptors.
  • The four most frequent post-processed terms are polite 1,377 (6.98%), talkative 1,341 (6.80%), friendly 1,281 (6.49%) and kind 1,264 (6.41%).
  • KMO is 0.896, but Kaiser suggests 34 factors and the scree four or six; eight is chosen as an intermediate value for interpretation.
  • The eight-factor solution explains 38.2%; loadings range from -0.64 to 0.78 and eight adjectives remain unassigned due to |loading|<0.25.
  • The authorial labels are decency, profoundness, instability, vibrancy, engagement, neuroticism, serviceability and subservience.
  • The largest moderate correlations include neuroticism-vibrancy 0.46, decency-neuroticism 0.39 and serviceability-vibrancy -0.38.
  • The unrestricted alpha is 0.57 for overlap=3 (23 pairs), 0.50 for overlap=2, 0.34 for overlap=1 and 0.31 with no shared conversations.
  • Alpha=0.78 requires overlap=3, same last conversation and only nine pairs.
  • The clearest Big Five overlap associates decency-agreeableness, neuroticism-neuroticism and serviceability-conscientiousness.
  • A psychologist manually assigned 95 of the 147 descriptors to Big Five and 24 remained without correspondence.
  • Against Völkel et al., the overlaps are 41% decency/approachable, 40% vibrancy/social-entertaining and 27% serviceability/serviceable.
  • The footnote of Figure 1 says 451 participants, but figure, abstract, method and 556-131 indicate 425.
  • The supplement only contains tables of descriptors/loadings/Big Five coding and top 100 frequencies.
  • No data or code release was located on the official surfaces or through title/DOI searches.

Limitations

  • The observed unit is human perception of co-produced conversations, not intrinsic personality of the model.
  • Personas, injected emotions, starts, user messages and content are confounded with text-davinci-002.
  • Each rater summarizes three distinct conversations into a single vector; there is no independent profile per conversation.
  • The matrix has 425 cases for 147 variables and heterogeneous stimulus sets.
  • Eight factors is not the direct recommendation of Kaiser or scree; the selection is post hoc.
  • Only 38.2% is explained and there is no parallel analysis, MAP, CFA, bootstrap or replication.
  • The factor labels are subjective interpretations.
  • The lexical pipeline includes manual decisions, removal of negations and screening with previous lists.
  • The clustering is greedy and order-dependent; t=10 is fixed by a visual elbow.
  • 23.5% of the survey is removed with style and time rules without public sensitivity.
  • Alpha=0.78 is based on nine pairs and an additional order condition.
  • Pairs without common conversations do not estimate reliability of the same object.
  • The Big Five overlap depends on manual assignment and does not use a validated measure on the same stimuli.
  • The comparison with voice assistants crosses studies and non-equivalent vocabularies.
  • Independence by person/start relies on descriptions and plots, not equivalence tests.
  • The sample is dominated by students; only 8% are native English speakers.
  • Interaction is incentivized with payments, badges, leaderboard and a CHF1,000 lottery.
  • Ratings are based on three short fictional conversations.
  • Only historical text-davinci-002 is evaluated and current generalization is uncertain.
  • Conversations, ratings, exclusions, clusters, translations or cleaning decisions are not public.
  • Code, dependencies, seeds, randomization or analysis pipeline are not public.
  • The appendix does not reproduce the complete rating survey instrument.
  • The figure 451 in the footnote of Figure 1 contradicts the analyzed n=425.
  • There is no preregistration, immutable release or independent replication.

What the study does not establish

  • It does not demonstrate that GPT-3 or a current LLM possesses personality.
  • It does not validate eight factors as a stable or universal taxonomy of chatbot personality.
  • It does not demonstrate that Big Five is psychometrically inadequate for all LLM agents.
  • It does not demonstrate that the factors replicate in another sample, language, culture or model.
  • It does not separate the effect of the model from the prompt, user, content, bundle or rater.
  • It does not demonstrate independence from persona or start by showing similar distributions.
  • It does not demonstrate a general reliability of 0.78; that figure comes from nine conditioned pairs.
  • It does not provide Big Five convergent validity through coded lexical overlap.
  • It does not automatically generalize to ChatGPT, GPT-4 or current systems.
  • It does not demonstrate that GPT-3 is more human than voice assistants through a comparison between studies.
  • It does not demonstrate a social transition from tool to integral companion.
  • It does not validate the use of the factors to control or compare deployed chatbots.
  • It does not allow numerical reproduction of results, tables and figures with public artifacts.
  • It does not generalize beyond mostly university participants and brief conversations in English.

Traceability

Scope: Full text

Version: ACM IMWUT 8(2), Article 61; author-lab copy dated May 2024, checked against official CGL copy dated June 2024

Consulted source: https://static.siplab.org/papers/imwut2024-personality_dimensions_of_gpt.pdf

Review: Codex full-text, visual, psychometric and artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-3 text-davinci-002 (October 2022 API version)
  • factor_analyzer exploratory factor analysis implementation

Instruments and metrics

  • Free-text three-adjective chatbot personality self-reports
  • 147-descriptor four-point perceived-personality rating inventory
  • Exploratory factor analysis with oblimin rotation
  • Kaiser-Meyer-Olkin sampling adequacy
  • Kaiser criterion and scree test
  • ten Berge factor-score projection
  • Pairwise Krippendorff alpha grouped by conversation overlap
  • Human-coded lexical correspondence with Big Five adjective lists

Data used

  • 1,998 human-text-davinci-002 conversations collected 7 October-6 November 2022 (not publicly released)
  • 9,267 three-adjective self-reports / 27,801 adjective entries (not publicly released)
  • 120 sampled conversation transcripts used by the rating survey (not publicly released)
  • 425 retained 147-descriptor rating vectors (not publicly released)
  • Official 10-page CGL supplemental tables

Evidence and location

  • Design, sample, model, prompts and parameters: Main paper pp. 5-8, sections 3.1-3.3 and Tables 1-2
  • Cleaning, translation, negations, clustering and t=10: Main paper pp. 8-10, section 3.4, Algorithm 1 and Figure 3
  • Survey, exclusions and three conversations per rater: Main paper pp. 10-13, sections 4.1-4.4 and Figures 4-5
  • KMO, factor retention and choice 5/8/10: Main paper pp. 13-14, section 4.5 and Figure 6
  • Variance, loadings, factors, correlations and scores: Main paper pp. 14-16, sections 5.1-5.2 and Tables 3-4
  • Big Five assignment and lexical comparison: Main paper pp. 15-18, sections 5.3-5.4 and Figure 8
  • Conditional alpha and N sizes: Main paper pp. 32-33, Appendix B and Table B.1
  • Limitations and generalization: Main paper pp. 18-23, sections 6.2-6.8 and Conclusion
  • 147 descriptors, loadings, coding and top 100: Official CGL supplement SHA-256 b38eb56596f78511cc45a5d4d0ee9c663bc99c179929dbdab4332c65c8cf6f3d, all 10 pages inspected
  • Versions and metadata: SIPLAB PDF SHA-256 7cb1ec4e9b365fe93f6af5395dd6a784c048fdf62936e86c0c7cfd8ca38d72f2; CGL PDF SHA-256 cc086cc97081e3c0efa2acdfe4e94e60cc5720a3b57114c19a02d82527588df4; Crossref DOI 10.1145/3659626
  • Absence of public data and code: SIPLAB, CGL, ETH Research Collection item 4eaf6ae0-e265-420d-bb5a-2f3c4b318c4b and GitHub title/DOI searches audited 16 Jul 2026
  • Complete report: reports/verification/article-211-gpt3-perceived-personality-factor-and-artifact-audit.json