<![CDATA[Generative artificial intelligence (GenAI) has become embedded in university writing practices, yet perception-based research cannot show what learners actually do when AI suggestions enter the composing process. This qualitative multiple-case study examined how 16 EFL undergraduates used, evaluated, accepted, rejected, and transformed GenAI assistance across four English-writing tasks completed over six weeks, with particular attention to learner agency, authorship, and academic integrity. Participants represented high-proficiency (n = 5), intermediate-proficiency (n = 7), and low-intermediate-proficiency (n = 4) writing profiles. The multimodal corpus comprised 64 task-specific final texts, 48.5 hours of screen recordings, and 342 recorded prompt-response decision episodes, complemented by interviews and stimulated-recall accounts. Episode-level coding identified four patterns: selective scaffolding (132 episodes; 38.6%), iterative co-composition (100; 29.2%), surface repair (73; 21.3%), and delegated composition (37; 10.8%). The first two patterns showed the strongest learner agency because students bounded AI use, modified prompts, challenged inaccurate outputs, and reworked generated material. Delegated composition represented the clearest loss of authorship, with whole AI-generated paragraphs copied with little verification or substantive revision. A stimulated-recall episode illustrates how ownership was preserved through rejection and restructuring of AI output. The findings show that responsible GenAI use is better understood as a sequence of evaluative decisions than as a binary distinction between AI use and non-use. The study proposes an AI-mediated writing agency framework centred on purpose, verification, transformation, disclosure, and responsibility.]]>
Copyrights © 2026
