Designing a scaffolded GenAI revision environment for EMI academic writing: a randomized active-control study of no-ai transfer and feedback use

Generative artificial intelligence (GenAI) can improve assisted drafts without demonstrating students’ ability to write independently. This randomized active-control study compared standard GenAI with a scaffolded revision environment integrating dialogic prompting, bounded Mandarin clarification, and a reflective revision loop in a Taiwanese English-medium instruction (EMI) writing module. The complete-case sample comprised 428 undergraduates. One week after revision, the scaffolded group achieved a higher score on a writing task without AI (75.86 vs. 72.99). Observed-score path analysis using standardized variables and a 0/1 condition indicator showed that the scaffolded condition was associated with post-task need support (β = .393, p < .001). Need support was associated with a higher coder-derived feedback-use index (β = .376, p < .001), which was associated with higher no-AI transfer-writing scores (β = .324, p < .001). Condition retained a direct association with transfer writing (β = .203, p = .007), while the theory-ordered serial indirect association was .048 (95% bootstrap CI [.026, .076]). Because need support was assessed after revision, the indirect pathway is interpreted as a model-based association rather than causal mediation. Thus, the randomized difference in writing after AI removal provides the strongest evidence, while the process model offers a provisional explanation.