Key points are not available for this paper at this time.
What happens when students write more fluently but engage less deeply in revision? This mixed-methods study examined how structured ChatGPT integration was associated with academic writing development among 320 Bangladeshi undergraduates in English for Academic Purposes courses. AI-assisted writing was associated with improved surface-level fluency and reduced writing anxiety, but also with lower authorial agency, weaker genre performance, and less recursive revision. To assess these shifts, the study introduces two analytic measures: revision entropy, capturing the variety and depth of revision moves, and the epistemic ownership index, estimating observable author control across writing decisions. Quantitative findings showed that AI-assisted drafts demonstrated polished language yet scored lower on coherence, thesis development, and source integration than drafts produced within teacher-scaffolded revision cycles. Revision entropy and epistemic ownership were significantly lower in the AI-mediated group. Qualitative data revealed four recurring tensions: compositional passivity, authorship uncertainty, genre detachment, and feedback fatigue. Conducted in intact (pre-existing, non-randomly assigned) classroom groups, the study reflects associations rather than causal effects. The findings underscore the need to structure AI use so it supports revision and rhetorical decision-making rather than replacing them.
Rahman et al. (Mon,) studied this question.