Preservice Special Education Teachers’ Acceptance of an AI-Based Microteaching Evaluation and Feedback System: A Pre–Post UTAUT Mixed-Methods Study

We examined preservice special education teachers’ acceptance of an AI system that supports the evaluation of, and feedback on, their own teaching demonstrations. The system analyzes recorded demonstrations against a structured rubric with time-stamped behavioral evidence, allows users to review and revise the AI evaluation through chatbot dialogue, and then generates personalized feedback. We used a convergent mixed-methods design informed by the Unified Theory of Acceptance and Use of Technology (UTAUT) and interpreted through an expectation-confirmation lens. Undergraduate students at a private university in South Korea completed a 23-item UTAUT-based questionnaire before and after a semester of use of the system in two required courses (valid pre-use responses N = 42; posttest N = 65; matched pairs n = 33), and 20 students were interviewed across the semester’s two demonstration cycles. Acceptance was uniformly high before use (composite means 4.22–4.63 on a 5-point scale) and did not decline after a semester of use; post hoc, exploratory one-sided non-inferiority tests rejected a decline of d = −0.35 or larger for every factor (all p ≤ 0.007). Performance expectancy increased modestly (d = 0.36; t-test p = 0.047, Wilcoxon p = 0.055, Holm-adjusted p = 0.282), a change that is best read as suggestive rather than established. At posttest, a regression of intention on the four UTAUT factors accounted for 62% of the variance, but effort expectancy (β = 0.37) showed the most consistent evidence of association with intention across classical, heteroscedasticity-robust, bootstrap, and influence-adjusted specifications; the association for social influence (β = 0.39 under classical estimation) was not robust to the estimator, performance expectancy was significant only when one influential case was excluded, and facilitating conditions were nonsignificant throughout. Facilitating conditions were, as the original model specifies, strongly associated with use behavior (β = 0.65) rather than with intention. In interviews, participants described their sustained acceptance in terms of evidence-grounded, time-stamped feedback, score discrepancies that prompted questioning and reflection, uptake into lesson planning and re-teaching, and the system’s role as an accessible complement to scarce instructor feedback, and they located the limits of that acceptance in context-recognition errors, feedback inconsistency, and interface friction. Taken together, the findings suggest that high pre-use expectations of AI-supported evaluation need not erode over a semester of sustained use when the system continuously displays the grounds of its judgments, and that ease of use, rather than perceived usefulness, is the factor showing the most consistent evidence of association with continued-use intention.