Attitudes towards AI: measurement and associations with personality

Evaluation and psychometric validity2024Scientific ReportsApproved editorial review

Authors: Jan-Philipp Stein, Tanja Messingschlager, Timo Gnambs, Fabian Hutmacher, Markus Appel

Keywords: Artificial Intelligence attitudes, Big Five personality, Dark Triad, Conspiracy beliefs, ATTARI-12 questionnaire

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

5
Authors
24
Findings
72
Limitations
18
Evidence

Editorial summary

English

The article develops ATTARI-12, a twelve-item scale intended to measure people's general attitude toward artificial intelligence, and examines whether that attitude is associated with personality traits. The scale balances six positively and six negatively worded items and represents cognitive, affective, and behavioral content, although it is scored as one general factor. The paper reports three human studies: 601 US MTurk participants for scale construction and validation, of whom 490 are retained; German students for translation and retest, with 166 responses at T1, 163 at T2, and 150 matched; and 353 MTurk participants for the personality analysis, of whom 298 are retained under the final attention criterion. No AI system's personality is measured.

In Study 1, the strict one-factor confirmatory model fits poorly, CFI=.85, RMSEA=.15, and SRMR=.08. A bifactor S-1 model adding an orthogonal factor for negative wording improves fit to CFI=.98, RMSEA=.06, and SRMR=.03. The general factor accounts for 79% of common variance and has hierarchical omega .83; the method factor accounts for 21% and has hierarchical omega .38. The total score has alpha .93 and correlates .60 with attitudes toward voice assistants, .68 with attitudes toward robots, and .04 with social desirability. The shared data reproduce these values. This supports a general score with a material wording effect; it does not show that the simple one-factor model fits well or establish comprehensive independent validity.

In Study 2, the German version has alpha .91 at T1 and .89 at T2 and a test-retest correlation of r=.804 after 26–36 days. AI attitude correlates with stated interest in an AI-related career, r=.635 at T1 and r=.563 at T2. The released OSF file contains only the 150 matched cases, so it reproduces the longitudinal correlations but cannot reconstruct from that file the 166 T1 and 163 T2 cases used in the table. The paper describes no translation and cultural-adaptation procedure and tests neither factor structure in German nor measurement invariance across English and German.

In Study 3, a cross-sectional hierarchical regression relates ATTARI-12 to age, gender, Big Five, Dark Triad, and conspiracy mentality. The final regression uses 297 complete cases, although the table labels it N=298, and explains 14% of the variance. Significant adjusted associations are younger age, beta=-.17, higher agreeableness, beta=.27, and lower conspiracy mentality, beta=-.22. Openness is p=.065; gender, conscientiousness, extraversion, neuroticism, and all three dark traits are nonsignificant. A preregistered minimum-time cutoff was changed from four to three minutes, increasing the retained sample from 247 to 298; the strict supplementary analysis retains the three main associations but reduces the Big Five increment and the agreeableness coefficient.

The defensible contribution is an open, reasonably stable scale for studying broad human attitudes toward AI, accompanied by auditable questionnaires, data, and syntax. Important limits remain: initial validation and CFA use the same sample; items were screened by the authors without cognitive interviews or external content validation; negative wording creates method multidimensionality; convergent evidence relies on brief measures from the same domain; and the personality study is correlational, self-reported, and based on convenience samples. The prose also reports RMSEA=.03 where Table 2 reports .06, and the regression caption says N=298 even though its degrees of freedom and the open data show 297 analyzable cases. Therefore, “predicts” means statistical association within this sample, not cause, a universal stable disposition, real adoption behavior, or attitudes toward any particular AI application.

Español

El artículo desarrolla ATTARI-12, una escala de doce ítems para medir la actitud general de personas hacia la inteligencia artificial, y estudia si esa actitud se asocia con rasgos de personalidad. La escala equilibra seis ítems positivos y seis negativos y representa contenidos cognitivos, afectivos y conductuales, aunque se puntúa como un único factor. El trabajo reúne tres estudios humanos: 601 participantes de MTurk en Estados Unidos para construir y validar la medida, de los que se retienen 490; estudiantes alemanes para traducción y retest, con 166 respuestas en T1, 163 en T2 y 150 emparejadas; y 353 participantes de MTurk para el análisis de personalidad, de los que se retienen 298 con el criterio final de atención. No se mide la personalidad de ningún sistema de IA.

En el Estudio 1, el modelo confirmatorio estrictamente unifactorial ajusta mal, CFI=0,85, RMSEA=0,15 y SRMR=0,08. Un bifactor S-1 que añade un factor ortogonal para la redacción negativa mejora a CFI=0,98, RMSEA=0,06 y SRMR=0,03. El factor general explica el 79% de la varianza común y alcanza omega jerárquica 0,83; el factor de método explica el 21% y tiene omega jerárquica 0,38. La puntuación total presenta alfa 0,93 y correlaciones de 0,60 con actitudes hacia asistentes de voz, 0,68 con actitudes hacia robots y 0,04 con deseabilidad social. Los datos compartidos permiten reproducir estas cifras. La evidencia apoya una puntuación general con un efecto de redacción relevante; no respalda que el modelo unifactorial simple ajuste bien ni una validez completa independiente.

En el Estudio 2, la versión alemana tiene alfa 0,91 en T1 y 0,89 en T2 y una correlación test-retest de r=0,804 tras 26–36 días. La actitud hacia IA se relaciona con el interés declarado por una carrera que incluya IA, r=0,635 en T1 y r=0,563 en T2. El archivo OSF publicado contiene solo los 150 casos emparejados, por lo que reproduce las correlaciones longitudinales pero no permite reconstruir desde ese archivo los 166 casos de T1 ni los 163 de T2 que usa la tabla. El artículo no describe un procedimiento de traducción y adaptación cultural ni prueba la estructura factorial o la invariancia entre inglés y alemán.

En el Estudio 3, una regresión jerárquica transversal relaciona ATTARI-12 con edad, género, Big Five, Tríada Oscura y mentalidad conspirativa. El modelo final usa 297 casos completos, aunque la tabla lo rotula N=298, y explica el 14% de la varianza. Las asociaciones ajustadas significativas son menor edad, beta=-0,17, mayor amabilidad, beta=0,27, y menor mentalidad conspirativa, beta=-0,22. Apertura queda en p=0,065; género, responsabilidad, extraversión, neuroticismo y los tres rasgos oscuros no son significativos. El cambio pre-registrado del umbral mínimo de cuatro a tres minutos eleva la muestra de 247 a 298; el análisis estricto suplementario conserva las tres asociaciones principales, pero reduce el incremento atribuido a Big Five y la magnitud de amabilidad.

La contribución defendible es una escala abierta y razonablemente estable para estudiar actitudes humanas generales hacia la IA, acompañada de cuestionarios, datos y sintaxis auditables. Sus límites son importantes: la validación inicial y el CFA se realizan en la misma muestra; los ítems fueron filtrados por los propios autores sin entrevistas cognitivas ni validación externa de contenido; la redacción negativa genera multidimensionalidad de método; las evidencias convergentes son medidas breves del mismo dominio; y el estudio de personalidad es correlacional, autorreportado y de conveniencia. Además, el texto dice por error RMSEA=0,03 donde Table 2 informa 0,06, y la leyenda de la regresión dice N=298 pese a que sus grados de libertad y el archivo abierto muestran 297 casos analizables. Por tanto, «predice» significa asociación estadística dentro de esta muestra, no causa, rasgo estable universal, conducta real de adopción ni actitud ante una aplicación concreta de IA.

Research question

Can a brief scale reliably represent a general human attitude toward AI in English and German, and what associations do age, gender, the Big Five, the Dark Triad, and conspiratorial mindset maintain with that score in a US MTurk sample?

Method

Three questionnaire studies. Study 1 recruited 601 US MTurk workers, applied prespecified and time-dependent exclusions down to N=490, compared four robust CFAs (one-factor, bifactor by facets, bifactor by wording, and combined model), and assessed consistency, attitudes toward voice assistants and robots, and social desirability. Study 2 administered the German version twice to students, with a mean separation of 31.4 days, and related the scale to career interest in AI. Study 3 recruited 353 MTurk workers, retained 298 after modifying the preregistered time cutoff, and estimated a hierarchical regression with eleven predictors; the final model uses 297 complete cases. The audit read and rendered the 16 pages, checked tables, appendix, and supplement, downloaded all OSF materials, and reproduced descriptives, correlations, filters, and standardized coefficients from the shared SAV files.

Sample: Study 1: 601 US MTurk participants; 111 exclusions and N=490 final, 212 women, 273 men, and 5 in other options or no response, age 19–72, M=39.78. Study 2: social science students from the University of Würzburg; 166 at T1, 163 at T2, and 150 matched, 113 women, 36 men, and 1 other option, M=21.21. Study 3: 353 US MTurk participants; N=298 with the final cutoff of three minutes, 102 women, 195 men, and 1 other option, M=39.29; the regression excludes the non-dichotomous gender case and uses N=297. These are convenience samples, not representative of AI users or the general population.

Findings

  • ATTARI-12 contains twelve items: six positive, six negative, and four each for cognitive, affective, and behavioral content.
  • The strictly one-factor CFA of Study 1 fits poorly: CFI 0.85, RMSEA 0.15, and SRMR 0.08.
  • The S-1 bifactor model of content facets also does not offer good fit: CFI 0.89, RMSEA 0.14, and SRMR 0.07.
  • The model with an orthogonal negative wording factor fits CFI 0.98, RMSEA 0.06, and SRMR 0.03.
  • The general factor explains 79% and the wording factor 21% of the common variance.
  • The hierarchical omega is 0.83 for the general factor and 0.38 for the wording factor.
  • Standardized loadings on the general factor range from 0.48 to 0.88.
  • The internal consistency of ATTARI-12 in Study 1 is alpha 0.93.
  • ATTARI-12 correlates 0.599 with attitudes toward voice assistants and 0.676 with attitudes toward robots.
  • The correlation with social desirability is 0.037 and not significant.
  • The German version reaches alpha 0.91 at T1 and 0.89 at T2.
  • Test-retest stability after 26–36 days is r=0.804.
  • Interest in an AI-related career correlates 0.635 with ATTARI-12 at T1 and 0.563 with ATTARI-12 at T2.
  • In Study 3, ATTARI-12 again shows alpha 0.92.
  • The mean attitude toward AI in Study 3 is 3.60 out of 5, slightly above the midpoint.
  • The final model with eleven predictors explains R²=0.14 of the variance in ATTARI-12.
  • Age is negatively associated with ATTARI-12 in the final model, beta=-0.17.
  • Agreeableness is the only significant Big Five trait in the final model, beta=0.27.
  • Conspiratorial mindset is negatively associated with ATTARI-12, beta=-0.22.
  • Openness falls below the conventional threshold, beta=0.11 and p=0.065.
  • Gender, conscientiousness, extraversion, neuroticism, Machiavellianism, psychopathy, and narcissism are not significant in the final model.
  • The strict analysis with a four-minute cutoff and N=247 retains age, agreeableness, and conspiratorial mindset, with smaller coefficients for the latter two.
  • Descriptives, correlations, inclusion filters, R², and main betas are reproduced from the OSF data.
  • The materials, data, and syntax of the three studies are publicly available on OSF.

Limitations

  • The scale is initially constructed and validated in a single sample.
  • There is no independent sample to confirm the structure selected in Study 1.
  • The final size of Study 1, N=490, falls below the minimum of 500 that the power analysis itself considered necessary.
  • The 120-second cutoff of Study 1 was set by inspecting the distribution after collecting the data.
  • The supplement acknowledges that the preregistration referred to an outlier analysis different from the cutoff ultimately applied.
  • Items were generated and discarded through internal discussion among the authors.
  • No content evaluation by external experts is reported.
  • No cognitive interviews with participants were conducted.
  • No qualitative pilot testing to check how the definition of AI is understood is reported.
  • The prior definition presents AI as capable of perceiving, acting, learning, and adapting autonomously and similarly to humans.
  • That definition may activate a concrete anthropomorphic representation of the attitudinal object.
  • The scale mixes statements about global benefits, emotions, and preference for use under a single score.
  • The strictly one-factor model shows clearly insufficient fit.
  • The unidimensional interpretation depends on modeling the negative wording effect separately.
  • The wording factor explains a non-trivial 21% of the common variance.
  • The article does not separately validate the cognitive, affective, and behavioral facets.
  • The CFA treats five-point ordinal responses with robust maximum likelihood instead of an explicit ordinal estimator.
  • No sensitivity of the results to other estimators or specifications is reported.
  • No cross-validation of the chosen bifactor model is published.
  • The convergent measures of Study 1 are brief and belong to the same technological domain.
  • The attitudes toward voice assistants scale was created ad hoc for this study.
  • The absence of correlation with a single social desirability scale offers limited discriminant evidence.
  • Incremental validity against existing AI attitude scales is not tested.
  • Predictive capacity is not compared with GAAIS, ATAI, AIAS, or other measures reviewed in the introduction.
  • Real behavior of use, adoption, avoidance, or purchase is not measured.
  • The career index of Study 2 is self-reported and conceptually close to the behavioral items of ATTARI-12.
  • Study 2 uses a small sample of students from a single program and university.
  • The German sample is young and predominantly female.
  • No direct translation, back-translation, bilingual committee, or cultural adaptation is described.
  • A CFA of the German version is not estimated.
  • Measurement invariance between English and German is not tested.
  • Invariance by age, gender, education, or culture is not tested.
  • The German social desirability measure was removed due to alpha below 0.40.
  • Study 2 is not listed as preregistered.
  • The OSF file for Study 2 contains only 150 matched cases.
  • That file does not allow reproducing the marginal descriptives of N=166 at T1 and N=163 at T2.
  • The retest of around one month does not establish stability over long periods or technological changes.
  • Studies 1 and 3 use US MTurk convenience samples.
  • The authors acknowledge that there may be repeated participants between Studies 1 and 3.
  • A possible overlap between the two MTurk samples is not identified or eliminated.
  • Study 3 has approximately two-thirds men.
  • Study 3 presents a relatively high educational level.
  • The sample is not weighted to represent the US population.
  • Prior experience, literacy, effective use, or technical knowledge of AI is not measured in the regression.
  • Media exposure, occupation, income, or ideology are not included as covariates in the main model.
  • The personality design is cross-sectional and does not establish temporal precedence.
  • Predictors and outcome are collected via self-report in the same session.
  • Common method variance may contribute to the associations.
  • The predictor term is statistical and not causal in this design.
  • The final model explains only 14% of the variance.
  • 86% of the variance in attitude remains unexplained by the included variables.
  • Openness with p=0.065 does not constitute conventional support for H1.
  • Eight of the eleven directional personality and demographic hypotheses do not receive support in the final model.
  • No explicit correction for the eleven hypotheses or the numerous coefficients examined is applied.
  • The change from a four-minute to a three-minute cutoff was decided after observing how many participants would be excluded.
  • That change increases the main sample from 247 to 298 participants.
  • The strict analysis reduces the R² increment of Big Five from 0.08 to 0.04.
  • The agreeableness coefficient drops from 0.27 to 0.19 in the strict analysis.
  • The conspiratorial mindset coefficient drops from -0.22 to -0.19 in the strict analysis.
  • Table 7 labels N=298, but F(11,285) implies N=297 and the file has one case without dichotomous gender.
  • The listwise exclusion of the non-binary gender case is not explained in the results text.
  • The text reports RMSEA=0.03 for the wording model, while Table 2 reports RMSEA=0.06.
  • Table 2 prints an apparently incomplete BIC for the facet model.
  • The claim of good validity is broader than the tests of convergence, social desirability, and retest presented.
  • Criterion validity with external decisions or outcomes is not evaluated.
  • Sensitivity to change after positive or negative experiences with AI is not evaluated.
  • A general attitude may conceal radical differences between medical AI, recommenders, robots, generators, and surveillance systems.
  • The observed associations may change as public salience and AI capabilities change.
  • Recommendations about transparency and fostering trust are not experimentally tested.
  • Mechanisms explaining why agreeableness or conspiratorial mindset relate to the score are not evaluated.
  • No independent replication by another team is conducted.
  • Relevance to synthetic personality is indirect: it studies human personality and attitudes toward AI, not traits of artificial agents.

What the study does not establish

  • It does not demonstrate that an attitude toward all AI is truly independent of context or application.
  • It does not demonstrate that the simple one-factor model of ATTARI-12 fits well.
  • It does not establish invariance between the English and German versions.
  • It does not separately validate cognitive, affective, and behavioral facets.
  • It does not demonstrate predictive validity for real adoption or avoidance behavior.
  • It does not demonstrate that personality causes attitudes toward AI.
  • It does not demonstrate that changing agreeableness or conspiratorial mindset changes attitude toward AI.
  • It does not demonstrate that chronological age is the mechanism responsible for more negative attitudes.
  • It does not support a significant association of openness, gender, conscientiousness, extraversion, neuroticism, or Dark Triad in the final model.
  • It does not allow generalizing the magnitudes to other countries, cultures, or representative samples.
  • It does not establish stability of ATTARI-12 over years or major changes in the AI ecosystem.
  • It does not demonstrate that ATTARI-12 outperforms previous AI attitude instruments.
  • It does not test that increasing transparency improves trust or acceptance.
  • It does not measure personality, awareness, emotion, or attitude of any AI system.
  • It does not directly inform how to design or validate a synthetic personality.

Traceability

Scope: Full text

Version: Scientific Reports 14:2909 (published 5 Feb 2024); DOI 10.1038/s41598-024-53335-2; CC BY 4.0

Consulted source: https://www.nature.com/articles/s41598-024-53335-2

Review: Codex full-text, visual, psychometric, preregistration, open-data, reproducibility, statistical-integrity and claim-scope audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Instruments and metrics

  • ATTARI-12: 12 items, five response points, six positive and six reverse-scored negative items
  • ATTARI-12 cognitive, affective and behavioral content facets
  • Three-item semantic differential for attitudes toward personal voice assistants
  • Three-item attitudes toward robots measure
  • 16-item Social Desirability Scale
  • Four-item AI-related career-interest index
  • 44-item Big Five Inventory
  • 27-item Short Dark Triad scale
  • Five-item Conspiracy Mentality Questionnaire
  • Confirmatory factor analysis with robust maximum likelihood
  • Bifactor S-1 models for content and negative-wording effects
  • Cronbach's alpha and hierarchical omega
  • Pearson correlations and hierarchical linear regression

Data used

  • Study 1 US MTurk all-participant SPSS dataset, N=601
  • Study 1 included-participant SPSS dataset, N=490
  • Study 2 matched German student SPSS dataset, N=150
  • Study 3 US MTurk SPSS dataset, N=353 with filter variables yielding N=298 and N=247
  • Study 1 CFA R code and rendered results
  • Study 1, Study 2 T1/T2 and Study 3 complete questionnaires
  • Study 1, Study 2 and Study 3 SPSS analysis syntax
  • English and German ATTARI-12 one-page instruments
  • Official Scientific Reports supplementary DOCX with Supplements S1–S8

Evidence and location

  • Metadata, DOI, license, and editorial dates: Scientific Reports 14:2909, pp. 1 and 16; DOI 10.1038/s41598-024-53335-2; published 5 Feb 2024; CC BY 4.0
  • Objectives, eleven hypotheses, and construct scope: Paper, abstract and introduction, pp. 1–4
  • Item generation, preregistration, and Study 1 sample: Paper, Study 1 Method, p. 5; Supplement S1–S2
  • CFA, loadings, fit, and wording effect: Paper, Tables 2–3 and Results, pp. 6–7
  • RMSEA 0.03 versus 0.06 discrepancy: Paper, Study 1 Results prose versus Table 2, p. 6; visual integrity audit 15 Jul 2026
  • Consistency and convergent and divergent validity of Study 1: Paper, Table 4 and Results, p. 7
  • Study 2 design, sample, and retest: Paper, Study 2 and Table 5, p. 8; Supplement S4–S5 and S8
  • Study 3 samples, exclusions, and time cutoff change: Paper, Study 3 Method, pp. 8–9; Supplement S6–S7
  • Personality correlations and regression: Paper, Tables 6–7 and Results, pp. 9–11
  • Actual N of the regression and legend discrepancy: Paper, Table 7 caption N=298 and F(11,285), p. 10; OSF Study 3 data contain 297 complete binary-gender cases
  • Limitations and authors' recommendations: Paper, General discussion and Conclusion, pp. 11–13
  • The twelve items and their classification: Paper, Appendix, pp. 13–14; OSF English and German ATTARI-12 PDFs
  • Frozen supplements S1–S8: .cache/editorial-sources/article-078/supplements/audit/41598_2024_53335_MOESM1_ESM.docx; sha256 411349d4012a33ed3b1f00a766af47626ebc1ff48d679d0992e8a8c2ce964fae
  • Frozen open data and materials: .cache/editorial-sources/article-078/supplements/audit/osf/; 16 official OSF files including SAV data, R/SPSS syntax, questionnaires and rendered CFA output; checked 15 Jul 2026
  • Study 1 reproduction: OSF study1-included-data.sav, sha256 716c714cd5a169af0a2e4c08dac830f1da0945827e59dcc929cd325e1ccef8ee; N=490, alpha=.9288, r=.5986/.6756/.0372
  • Study 2 reproduction and incomplete coverage: OSF study2-data.sav, sha256 65433eb3b00c42966b184184eea7e18aa0a45431f22cd104dc420802cb2dc169; 150 matched rows only, r=.8035/.6275/.5628
  • Study 3 reproduction: OSF study3-data.sav, sha256 937aeaa61b2ed1648807f1d08169bde33ca9dedadaa4214f6bc80b4857079f58; filters N=298/N=247; complete-case N=297; R2=.1385; betas reproduced
  • Integral reading and visual check: All 16 PDF pages rendered and inspected, including Tables 1–7, Appendix, author contributions, data availability and references; checked 15 Jul 2026