Beyond Trust Recovery: HowReappraisal and Verification Shape Trajectories of Anxiety, Trust, and Calibrated Reliance After Generative AI Feedback Errors in Higher Education

Generative artificial intelligence feedback requires learners to distinguish useful recommendations from plausible errors. We examined whether reappraisal adds value to structured verification after a shared AI error. In an individually randomized 2 × 2 experiment, 480 Chinese undergraduates received neutral reflection or reappraisal and ordinary evidence access or structured verification. Six writing-decision blocks yielded 10,792 responses from 476 learners. Discrimination was the difference between acceptance probabilities for correct and incorrect advice. The marginal incremental effect on discrimination was 0.100 (95% bootstrap CI [0.042, 0.162]). The reconstructed original-aggregate anxiety contrast favored combined support by 0.210 scale points, whereas the trust–accuracy coupling contrast included zero. The baseline-adjusted independent-transfer contrast was 0.957 points on a 36-point scale (95% CI [−0.447, 2.362]). These findings distinguish correctness-sensitive reliance from trust recovery and independent performance within a fixed task bank.