Chinese music students’ trust in GenAI linked most strongly to intention to use it for composition

A cross-sectional survey of 892 undergraduate music students with previous hands-on experience using generative artificial intelligence for composition-related tasks found that trust in AI and attitudes toward its use were most strongly associated with behavioral intention to use GenAI in music composition. Perceived behavioral control and subjective norms had more limited associations in this predominantly low-frequency user group, and the extended model explained 56.2% of the variation in intention.

The study, published in Discover Artificial Intelligence on 12 October 2026, extended the Theory of Planned Behavior with task-technology fit, AI trust and AI anxiety. Participants were undergraduates specializing in vocal or instrumental music at universities in five provincial-level regions of China: Shandong, Inner Mongolia, Hubei, Zhejiang and Jiangsu. They were recruited through convenience and purposive sampling, peer referrals and voluntary participation in scheduled music classes.

Eligible students had to be undergraduate vocal or instrumental music majors who had previously used GenAI for composition-related activities. The questionnaire covered music or song generation, assisted composition and arrangement, and singing-voice synthesis, rather than a single tool or category. Of 1,230 returned questionnaires, 892 were retained after screening for excessively rapid online responses, near-invariant answers and multivariate outliers, amounting to 72.5% of those returned. The final sample was 75.7% female, with 52.0% studying vocal music and 48.0% instrumental music. Most participants, 73.3%, used GenAI for composition less than once a week.

The researchers used a seven-construct questionnaire covering behavioral intention, attitude, perceived behavioral control, subjective norms, task-technology fit, AI trust and AI anxiety. Translation and back-translation were reviewed by English professors and two music experts, while a separate pretest with 50 eligible participants produced Cronbach’s alpha values from 0.735 to 0.822. These pretest participants were not included in the final analysis. The main measurement model showed satisfactory internal consistency, indicator reliability and convergent validity, with Cronbach’s alpha values of 0.802 to 0.904, item loadings of 0.769 to 0.917 and average variance extracted values of 0.647 to 0.830.

Partial least squares structural equation modeling, assessed with 5,000 bootstrap resamples, estimated the proposed associations. Attitude had a positive direct association with behavioral intention, with a standardized coefficient of 0.264, while perceived behavioral control had a smaller positive association, at 0.102, and subjective norms had a still smaller association, at 0.057. Task-technology fit was positively associated with intention at 0.171. AI trust showed a larger positive direct association, at 0.314, and had a medium effect size according to the reported f² measure of 0.169. AI anxiety was negatively associated with intention, at −0.214.

The extended constructs were also associated with attitude and perceived behavioral control. Task-technology fit was positively associated with attitude at 0.317 and with perceived behavioral control at 0.118. AI trust was positively associated with attitude at 0.278 and perceived behavioral control at 0.112. AI anxiety was negatively associated with attitude at −0.109 and perceived behavioral control at −0.340. Subjective norms were positively associated with attitude at 0.131, but their association with perceived behavioral control was not statistically significant. The model also produced an acceptable approximate fit, with a standardized root mean square residual of 0.038.

A complementary artificial neural network analysis ranked the direct predictors by their relative importance for predicting behavioral intention. AI trust ranked first with normalized importance of 100%, followed by attitude at 70.22%, task-technology fit at 56.94%, AI anxiety at 48.80%, perceived behavioral control at 32.20% and subjective norms at 14.70%. These normalized values describe relative predictive importance within the trained network; they are not path coefficients, significance tests or measures of association direction. Across ten runs, mean root mean square error was 0.468 for training and 0.447 for testing, with no clear evidence of substantial overfitting.

The results suggest that trust in AI outputs and students’ evaluations of whether GenAI belongs in their creative practice may be especially relevant to future-use intention among this group. The authors propose that educators could help students form calibrated judgments by comparing, revising and justifying AI-generated musical material, assessing its suitability for particular tasks and retaining authority over final creative decisions. They also suggest that clear guidance on acceptable use, authorship and attribution may help address uncertainty, while structured opportunities to reject or override AI-generated material may support students’ sense of agency.

These findings are limited by the cross-sectional design, which cannot establish temporal order or causal effects. Convenience and purposive sampling may restrict generalizability, and the predominance of low-frequency users means the results primarily reflect students with limited experience of GenAI-assisted composition. The study also examined general intention rather than actual use, completed compositions or behavior at particular stages of the creative process. It was not designed to compare specific tools, task categories or compositional stages, and its findings should therefore be applied cautiously beyond similar Chinese music-student populations.

Subject of Research: Cross-sectional study of Chinese music students’ intention to use generative AI for music composition

Article Title: An extended theory of planned behavior model of generative AI use intention for music composition among Chinese music students

Article References: Pan, L., Chen, Q., & Liu, Y. (2026). An extended theory of planned behavior model of generative AI use intention for music composition among Chinese music students. Discover Artificial Intelligence, 6(1), Article 1441. https://doi.org/10.1007/s44163-026-02460-0

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02460-0

Keywords: generative AI, music composition, music students, China, AI trust, AI anxiety, task-technology fit, Theory of Planned Behavior, behavioral intention, extended, theory, planned