Generative AI Acceptance in Higher Education: A Meta-Analysis and Meta-Analytic Structural Equation Model of the 2026 PLS-SEM Literature

Hundreds of survey studies have modelled the acceptance of generative artificial intelligence (GenAI) in higher education with partial least squares structural equation modeling (PLS-SEM), but their path coefficients have not been pooled, and it is not known whether the determinants established for earlier technologies hold, or hold evenly, for GenAI. We address this gap with the first meta-analysis of the 2026 PLS-SEM literature. Following PRISMA 2020, a Scopus search returned 151 records, of which 119 met strict eligibility. Standardized direct path coefficients and analytic sample sizes were extracted and double-checked by hand from each study's structural-model table. Six relationships with adequate coverage were pooled with random-effects models on Fisher-z-transformed coefficients (DerSimonian-Laird), and we report confidence and prediction intervals, subgroup and meta-regression moderators, study-level risk-of-bias appraisal, leave-one-out sensitivity analysis, and Egger's test with trim-and-fill. With between 11 and 20 studies per relationship and a combined sample of up to 10,819, all pooled effects were positive and significant: performance expectancy or usefulness predicted behavioural intention at beta = 0.33 (95% CI 0.25-0.41), effort expectancy or ease of use at 0.21, social influence at 0.21, and attitude at 0.58; behavioural intention predicted use at 0.52. Heterogeneity was very high throughout (I-squared 80-99%), and the prediction intervals were wide enough to include zero for four of the six relationships. A coherent regional pattern emerged: usefulness and effort expectancy were both weaker in Chinese samples (0.27 and 0.09) than elsewhere (0.37 and 0.29), whereas social influence was stronger (0.25 versus 0.17). Pooled estimates were robust to the removal of any single study; trim-and-fill imputed no missing studies. The acceptance hierarchy familiar from TAM and UTAUT is reproduced, but the determinants are far less stable than the pooled means suggest, and a single coefficient is a poor guide to any new setting.