关键词: AI素养, 自我效能, 大学生, 贝叶斯元分析, 国际比较, 生成式人工智能

Abstract: Objective: Previous studies on the relationship between AI literacy and self-efficacy have reported inconsistent findings because of varied conceptualization, measurement approaches and sample contexts. Importantly, the relationship is theoretically complex. Students with stronger AI literacy may experience an enhanced sense of control and competence, while those with greater awareness of AI’s capabilities may potentially erode confidence in their competence. To address this issue, the present study used a Bayesian meta-analysis to systematically synthesize the available evidence and identify the moderating conditions under which the association varies.
Method: This study followed PRISMA 2020 guidelines and systematically searched Web of Science (SCI/SSCI), Scopus, and CNKI through November 2025. A total of 36 independent studies involving 25,251 university students were included. A Bayesian random-effects model with weakly informative priors [Normal (0,1) for the pooled effect; Half-Normal (0.5) for heterogeneity] was used as the primary analytical framework. Model convergence was evaluated through MCMC sampling with R-hat and ESS convergence checks. Prior sensitivity analyses (Half-Normal (1.0)) confirmed robustness. Frequentist random-effects models were also conducted for robustness checks. Moderator analyses covered self-efficacy type, AI literacy measurement approach, student discipline, country/region, and study design. Publication bias was evaluated via funnel plot inspection and Egger's regression. Methodological quality was assessed using the Newcastle-Ottawa Scale (inter-rater ICC = 0.938).
Results: The Bayesian model yielded a robust positive overall association (posterior median r = 0.495, 95% CrI [0.420, 0.570]; frequentist r = 0.518), with substantial heterogeneity (I² = 91.73%, τ = 0.32). Publication bias was negligible (Egger's p = .358), and leave-one-out sensitivity analyses confirmed stability (r range: 0.50-0.53). Four significant moderators emerged (all Qb p < .05); country/region was not significant (Qb = 0.76). Self-efficacy type was the strongest moderator: associations were highest for creative self-efficacy (r = 0.71) and entrepreneurial self-efficacy (r = 0.61), exceeding those for AI-specific self-efficacy (r = 0.53) and general academic self-efficacy (r = 0.37). Studies using self-developed or integrated self-report scales yielded larger effects (r = 0.71) than those using objective AI literacy tests (r = 0.14). Students in education, media, and language disciplines showed stronger associations (r = 0.60) than those in engineering/computer science (r = 0.45) or business (r = 0.41). Within-region heterogeneity was equally high in East Asian (I² = 78.22%) and Western samples (I² = 77.52%), indicating that meaningful variation resides within rather than between geographic regions.
Contributions: This study makes three major contributions. First, it adopts a Bayesian meta-analytic framework. Compared with traditional frequentist approaches, it more effectively captures both effect and heterogeneity parameters. Second, the moderation pattern challenges domain-matching intuitions: AI literacy's strongest links are not with AI-specific self-efficacy nor with technically oriented students (engineering and computer science), but with creative and entrepreneurial efficacy beliefs and with students from humanities-adjacent disciplines. This counterintuitive result suggests AI literacy operates less as a narrow technical competency reinforcing task-specific confidence than as a transferable psychological resource activating higher-order efficacy beliefs around creativity, career adaptation, and future agency. In the generative AI era, this reframes AI literacy as a form of human capital with broad motivational transfer rather than a domain-bound skill. Third, although no significant cross-national differences were found, substantial within-region heterogeneity remained. This suggests that micro-level instructional factors, such as curriculum design, Al integration strategies, and teacher support, may play a more important role than broader cultural or policy contexts.
Conclusions:Overall, AI literacy is positively associated with university students’ self-efficacy, although the effect strengths depend substantially on self-efficacy type, measurement approach and disciplinary context. These findings suggest that AI literacy education should move beyond technical training alone and place greater emphasis on creativity, adaptability and higher-order efficacy beliefs. Higher education institutions should develop differentiated AI literacy curricula based on students’ disciplinary backgrounds and developmental needs. Future research should emphasize longitudinal and experimental design, as well as more standardized measurement approaches, in order to better identify the causal mechanism underlying the relationship between AI literacy and self-efficacy.

Key words: AI literacy, self-efficacy, university students, Bayesian meta-analysis, international comparison, generative AI