Explaining teacher adoption of AI through the technology acceptance model: a meta-analysis and systematic review

The integration of Artificial Intelligence (AI) in education has accelerated rapidly, yet the factors shaping teachers’ adoption of these technologies remain insufficiently understood. A growing body of empirical research applies the Technology Acceptance Model (TAM) to examine how teachers integrate AI, focusing on perceived usefulness (PU), perceived ease of use (PEOU), and behavioral intention (BI). TAM was, however, originally developed for static, deterministic technologies, and AI differs in its opacity, adaptivity, and capacity to challenge professional roles and identity, raising the question of whether perceived usefulness and ease of use alone are sufficient to predict adoption intentions in this context. Following PRISMA guidelines, this study conducts a systematic review of 38 empirical studies ( N = 15,919 teachers across 19 countries) on teachers’ AI adoption, together with a meta-analysis of the subset reporting extractable bivariate correlations. A random-effects model (DerSimonian–Laird) is used to estimate the associations among PEOU, PU, and BI across $$k = 24$$ studies ( N = 11,768 ), while the full corpus is synthesised to identify contextual conditions shaping these relationships. Pooled correlations are positive and moderate to strong (PEOU–PU: $$r = 0.532$$ , 95% CI [0.410, 0.636]; PEOU–BI: $$r = 0.536$$ , 95% CI [0.435, 0.623]; PU–BI: $$r = 0.619$$ , 95% CI [0.511, 0.708]), and are robust to alternative variance estimators and to the inclusion of UTAUT studies under a construct-equivalence ruling. Heterogeneity is very high ( I² = 97.9% – $$98.6%$$ ) and prediction intervals span from negative to near-unity values, indicating that these associations hold directionally but vary considerably across contexts in ways the core model does not represent. Egger’s regression indicates funnel asymmetry on all three associations; leave-one-out diagnostics attribute this asymmetry entirely to a single large-sample, low-effect study, and the asymmetry is not interpretable as evidence of publication bias under heterogeneity of this magnitude. The qualitative synthesis identifies five dimensions associated with variation in these relationships: AI literacy and trust as preconditions for usefulness judgments; a shift from functional to identity-based acceptance shaped by pedagogical beliefs and institutional norms; the role of sociocultural and institutional context; tool-specific adoption logics; and ethical risk calibration as a values-based gate on adoption decisions. These findings suggest that predicting teacher AI adoption intentions may require variables beyond the TAM core, and motivate the development of extended frameworks designed for the sociotechnical complexity of AI in education.