Integrating Generative Artificial Intelligence (GenAI) into traditional teacher feedback is recommended as a potential way to improve students' learning performance. However, empirical evidence remains limited on how GenAI-augmented teacher feedback affects students' Classical Chinese reading comprehension across three levels (literal, interpretative, and critical), particularly regarding its immediate and delayed effects. To address this gap, this study conducted a twelve-week experiment with 72 Chinese middle school students. The experimental group (N = 35) received GenAI-augmented teacher feedback, while the control group (N = 37) received traditional teacher feedback only. An analysis of covariance (ANCOVA), using the pre-test scores as a covariate, was conducted on the three-stage reading comprehension tests (pre-test, post-test, and delayed test), complemented by post-experimental interview analysis. The results demonstrated that students who received augmented feedback experienced an initial improvement in literal and interpretative comprehension that declined over time, though the decline was slower than in the control group. This suggests that GenAI may provide additional value to teacher feedback at the levels of literal and interpretative comprehension over the longer term. Moreover, students' critical comprehension improved in both post-test and delayed test, though the difference between the two groups was not significant, highlighting the limitations of GenAI in supplementing teacher feedback to support students' critical comprehension. Therefore, educators should leverage GenAI for lower-level reading comprehension to support teachers while prioritising teacher-led feedback for higher-level critical reading comprehension. This study provides a theoretical basis and practical reference for effectively integrating GenAI in education.
Who is mimicking whom when university EFL writers and ChatGPT interact over a sustained period? This 10-week longitudinal study tracked 20 first-year Chinese undergraduates across a Pre-AI baseline, an eight-week ChatGPT-assisted intervention, and a Post-AI independent phase, drawing on corpus analysis of 400 essays alongside 320 interaction logs and triangulated qualitative data. Twenty of twenty-two linguistic features showed large phase effects, but the more telling pattern was selective: features carrying semantic or pragmatic weight, including nominalisation density, hedging, and rebuttal quality, were retained well after AI withdrawal, whereas stylistically marked but optional structures such as cleft constructions and formal connectives reverted towards baseline. Students appear to have consolidated the resources that allowed them to mean differently rather than only to sound differently. Two further findings complicate received accounts: an AI-generated content (AIGC) paradox in which detection rates registered measurable signal even in pre-intervention writing and tracked closely with holistic ratings (r = .862), a pattern more readily read as evidence of formal academic register than of machine authorship; and a recursive coupling in which student prompts and AI responses progressively shape each other. The study proposes a Mutual Adaptation Framework that takes the learner-tool system, rather than the learner alone, as the developmental unit.
