Who is GPT-3? An exploration of personality, values and demographics

Evaluation and psychometric validity2022ACL AnthologyApproved editorial review

Original title: Who is GPT-3? An Exploration of Personality, Values and Demographics

Authors: Marilù Miotto, Nicola Rossberg, Bennett Kleinberg

Keywords: Language models, GPT-3, Psychological assessment, Personality traits, Computational social science

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

3
Authors
7
Findings
10
Limitations
8
Evidence

Editorial summary

English

This NLP+CSS 2022 workshop paper administers three kinds of self-report to the original GPT-3 DaVinci: HEXACO-60, the 21-item Human Values Scale (HVS), and open age and gender questions. Each item is queried separately at temperatures 0.0 through 1.0 in .1 steps; one output is retained at temperature zero and 100 are requested per item at every other temperature. Non-numeric outputs remove 1.73% of HEXACO and .004% of HVS responses. This structure is not a sample of people: without memory, different questionnaire items are independent completions, and any composite profile groups positions by index without a shared identity or trajectory. Multivariate tests treat those combinations as observations, making between-profile variance, facet covariance, and human comparison difficult to interpret psychometrically. For demographics, approximately 1,001 completions per question yield mean age 27.51 (SD 5.75; range 13–75), 66.73% female, 31.87% male, and 1.40% other outputs. Age falls 5.81 years per unit of temperature. The gender analysis is internally inconsistent: it names female vs not female as the dependent variable and reports β = +1.18, but interprets e^1.18 as increased male odds; the table does show a general, non-monotonic decline in female responses. These are sampling-conditioned text patterns, not model demographics. Aggregate HEXACO means are 3.75 Honesty-Humility, 3.05 Emotionality, 3.51 Extraversion, 3.18 Agreeableness, 3.54 Conscientiousness, and 3.59 Openness. Temperature is negatively associated with Emotionality (β = −.23) and positively associated with Extraversion (.31), Agreeableness (.40), Conscientiousness (.25), and Openness (.17), all p < .001; Honesty-Humility is not significant at p < .01. The discussion later reverses several directions: it treats lower Honesty-Humility as less willingness to manipulate, refers to higher Emotionality despite its negative coefficient, and says the remaining four facets decrease although the analysis says they increase. Facet correlations only partly match human references, and the results themselves say no consistent pattern emerges. Without memory, HVS Table 4 reports means from 4.51 to 5.95 on a six-point scale, contradicting the text's claim that all lie between four and five, and lower variance than references. Nine of ten values decline with temperature; Stimulation does not. With previous-response memory, tested only for HVS at temperatures 0, .2, .4, .6, .8, and 1, the pattern changes radically: total means are 2.17 Conformity, 2.74 Tradition, 5.30 Benevolence, 5.39 Universalism, 5.42 Self-Direction, 5.25 Stimulation, 4.01 Hedonism, 3.85 Achievement, 2.92 Power, and 2.69 Security. This shows that context format dominates the supposed profile. The authors interpret greater theoretical coherence but concede that raw human data would be needed for formal comparison; their inter-value correlations have little overlap with the human matrix. The text also calls the Security and Hedonism effects negative while printing positive β values of .21 and .88; the tables decline and suggest omitted minus signs. The historical contribution is evidence that temperature and prompt memory strongly alter questionnaire outputs. It does not establish that GPT-3 has personality, values, or demographics: there is no external behavior, construct validity, test–retest of one stable entity, inferential equivalence with humans, or control for instruments likely present in training data.

Español

Este artículo del workshop NLP+CSS 2022 administra al GPT-3 DaVinci original tres tipos de autoinforme: HEXACO-60, Human Values Scale (HVS) de 21 ítems y preguntas abiertas de edad y género. Cada ítem se consulta por separado con temperaturas de 0,0 a 1,0 en pasos de 0,1; a temperatura 0 se conserva una salida y en las demás se piden 100 por ítem. Se excluye 1,73 % de HEXACO y 0,004 % de HVS por respuestas no numéricas. Esta estructura no equivale a una muestra de personas: sin memoria, las respuestas de distintos ítems son completados independientes y cualquier perfil compuesto agrupa posiciones por índice sin una identidad o trayectoria compartida. Las pruebas multivariantes tratan esas combinaciones como observaciones, lo que hace que varianza entre perfiles, covarianzas entre facetas y comparación con humanos sean difíciles de interpretar psicométricamente. En demografía, 1.001 completados aproximados por pregunta producen edad media 27,51 (DE 5,75; rango 13–75), 66,73 % respuestas femeninas, 31,87 % masculinas y 1,40 % otras. La edad baja 5,81 años por unidad de temperatura. El análisis de género es internamente inconsistente: declara como variable dependiente female vs not female y β = +1,18, pero interpreta e^1,18 como aumento de odds masculinas; la tabla sí muestra una reducción general, aunque no monotónica, de respuestas femeninas. Son patrones de texto provocados por sampling, no demografía del modelo. En HEXACO, las medias agregadas son 3,75 en honestidad-humildad, 3,05 en emocionalidad, 3,51 en extraversión, 3,18 en amabilidad, 3,54 en responsabilidad y 3,59 en apertura. La temperatura se asocia negativamente con emocionalidad (β = −0,23) y positivamente con extraversión (0,31), amabilidad (0,40), responsabilidad (0,25) y apertura (0,17), todas p < 0,001; honestidad-humildad no es significativa al umbral p < 0,01. La discusión invierte después varias direcciones: atribuye a menor honestidad más rechazo a manipular, habla de mayor emocionalidad cuando el coeficiente es negativo y afirma que las otras cuatro facetas disminuyen cuando el análisis dice que aumentan. Las correlaciones entre facetas solo coinciden parcialmente con referencias humanas y el propio resultado dice que no aparece un patrón consistente. En HVS sin memoria, la Tabla 4 arroja medias de 4,51 a 5,95 sobre seis puntos, contradiciendo el texto que afirma que todas están entre 4 y 5, y una varianza menor que las referencias. Nueve de diez valores disminuyen con temperatura; stimulation no cambia. Con memoria de respuestas previas, probada solo en HVS a temperaturas 0,0, 0,2, 0,4, 0,6, 0,8 y 1,0, el patrón cambia radicalmente: las medias totales son 2,17 conformity, 2,74 tradition, 5,30 benevolence, 5,39 universalism, 5,42 self-direction, 5,25 stimulation, 4,01 hedonism, 3,85 achievement, 2,92 power y 2,69 security. Eso muestra que el formato de contexto domina el supuesto perfil. Los autores interpretan mayor coherencia teórica, pero admiten que haría falta comparar datos humanos crudos; sus correlaciones intervalor tienen poco solapamiento con la matriz humana. Además, el texto califica como negativos los efectos de security y hedonism aunque imprime β positivos de 0,21 y 0,88; las tablas descienden y sugieren signos omitidos. El aporte histórico es demostrar que temperatura y memoria del prompt alteran fuertemente salidas de cuestionarios. No demuestra que GPT-3 tenga personalidad, valores o demografía: no hay conducta externa, validez de constructo, test–retest de una entidad estable, equivalencia inferencial con humanos ni control de contaminación por instrumentos probablemente presentes en el entrenamiento.

Research question

What HEXACO profiles, HVS values, and demographic responses does GPT-3 DaVinci produce under different temperatures, and how much does the pattern change when the prompt retains previous responses?

Method

GPT-3 DaVinci is queried via API with HEXACO-60, HVS-21, and age/gender questions. Each item is sent separately at eleven temperatures; one output is requested at 0.0 and 100 at each level 0.1 to 1.0. Composites, descriptives, regressions, MANOVA, and correlations are computed, comparing them descriptively with published human studies. For HVS a cumulative design is repeated that includes previous items and responses at six temperatures.

Sample: A single model. Per item one response is obtained at temperature 0 and 100 at each of the other ten temperatures; these are independent completions, not participants. HEXACO contains 60 items, HVS 21, and the two demographic questions are administered separately. HVS with memory is run at six temperatures. No human participants or matched raw human data are collected.

Findings

  • Demographic outputs average 27.51 years and 66.73% female; age decreases with temperature, but the coefficient and the interpretation of the gender model are incompatible.
  • HEXACO means are H 3.75, E 3.05, X 3.51, A 3.18, C 3.54, and O 3.59; temperature reduces E and raises X, A, C, and O in the main analysis.
  • The subsequent discussion reverses the direction or meaning of several HEXACO effects, so it is not a coherent source for interpreting them psychologically.
  • Without memory, HVS reaches means 4.51 to 5.95, not 4 to 5 as the text claims, and nine values decrease as temperature increases.
  • The HEXACO correlation matrix does not reproduce a consistent human pattern; HVS correlations without memory are much lower than human ones.
  • Adding previous responses changes HVS to high benevolence, universalism, self-direction, and stimulation, and low conformity, tradition, power, and security.
  • Memory raises many correlations among values, but the matrix continues to have little overlap with the human reference and the required formal contrast is not performed.

Limitations

  • Each item without memory is generated independently; composing responses by position does not create stable subjects nor support psychological covariances among facets.
  • Temperature controls sampling randomness and simultaneously defines the analyzed groups; a change in token distribution does not equate to a change in personality.
  • A single GPT-3 engine is studied without an exact snapshot and with default parameters not detailed; there is no temporal or between-model replication.
  • The scales are human self-reports and it is not validated that their items, scoring, or factors measure equivalent constructs in a model.
  • Human comparisons use published means, deviations, and matrices from different samples, not raw data subjected to the same design or equivalence tests.
  • Lower variance than humans may result from prompting, repetition, or sampling and does not demonstrate psychometric consistency.
  • No alpha/omega, test-retest, invariance, convergent, discriminant, criterion, or behavioral validity is calculated.
  • The instruments may have been in the training data; the authors do not know GPT-3's data and acknowledge exposure or demand characteristics as an alternative explanation.
  • Memory is only tested on HVS and turns each previous response into a signal for subsequent ones, potentially imposing coherence through contextual continuation.
  • The article contains inconsistencies of sign, range, and narrative in gender, HEXACO, and HVS that prevent accepting several interpretations without reservations.

What the study does not establish

  • It does not demonstrate that GPT-3 is a person, has identity, age, gender, values, or psychological traits.
  • It does not demonstrate that a collection of independent completions represents one or multiple synthetic participants.
  • It does not establish equivalence between output profiles and human distributions nor that the model is a valid substitute in the social sciences.
  • It does not allow inferring honesty, anxiety, sociability, or behavior from self-report responses.
  • It does not demonstrate human coherence of HVS with memory; the paper itself calls for raw human data to test it.
  • It does not separate instrument learning, textual desirability, response bias, and instruction following from actual personality.
  • It does not generalize to current models, other languages, conversational prompts, or decisions outside the questionnaire.

Traceability

Scope: Full text

Version: NLP+CSS 2022 proceedings, ACL Anthology 2022.nlpcss-1.24

Consulted source: https://aclanthology.org/2022.nlpcss-1.24.pdf

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI GPT-3 DaVinci (175B; exact engine snapshot not reported)

Instruments and metrics

  • 60-item HEXACO personality inventory
  • 21-item Human Values Scale from the European Social Survey
  • Open-ended age and gender prompts
  • Multivariate analyses and per-facet regressions
  • Inter-facet and inter-value Pearson correlations
  • Cumulative response-memory prompting

Data used

  • GPT-3 item-level API completions across eleven temperatures
  • Ashton and Lee (2009) HEXACO college and community reference summaries
  • Schwartz et al. (2015) global and German Human Values Scale reference summaries
  • Response-memory HVS completions across six temperatures

Evidence and location

  • Objective, anthropomorphic scope, and initial caveats: NLP+CSS 2022, pp. 218 to 219, Abstract and Introduction
  • HEXACO, HVS, prompts, and temperature sampling: NLP+CSS 2022, pp. 219 to 220, sections 2.1 to 2.4 and Figure 1
  • Demographic results and inconsistency of the gender model: NLP+CSS 2022, p. 221, section 3.1 and Table 1
  • HEXACO means, temperature effects, and correlations: NLP+CSS 2022, pp. 221 to 222, section 3.2 and Tables 2 to 3
  • HVS without memory, range of means, and temperature effects: NLP+CSS 2022, pp. 222 to 223, section 3.3 and Tables 4 to 5
  • HVS with memory design and results: NLP+CSS 2022, pp. 223 to 224, section 3.4, Figure 2 and Tables 6 to 7
  • Contradictory interpretations and human comparison: NLP+CSS 2022, pp. 224 to 226, sections 4.1 to 4.1.2
  • Training contamination, absence of formal comparison, and limitation to a single model: NLP+CSS 2022, p. 226, section 4.2