<![CDATA[Research on AI-assisted writing in EFL contexts has been growing rapidly. However, most studies rely on self-reported perceptions, leaving a gap between what students say about their AI use and what they actually do. This qualitative multiple case study examined EFL students' perceptions of ChatGPT and their observable interaction patterns during the writing of a thesis proposal. The study involved three undergraduate EFL students from Universitas Islam Batang Hari. Data were gathered using semi-structured interviews with stimulated recall, full ChatGPT conversation logs shared through an export link, and a structured observation checklist. Interview data were analyzed thematically, while the documentary data 31 substantive prompts in total were analyzed using a six-category prompt classification taxonomy and a four-category editing behavior typology. Results for RQ1 revealed three perception themes: ChatGPT as cognitive scaffold, awareness of referential unreliability, and ambivalent writing confidence. Results for RQ2 indicated that generation prompts were the most common prompt type across all participants, with prompt sophistication associated with participants' experience level and language preference. Analysis of editing behavior indicated a predominant pattern of selective use and substantial modification, with direct adoption occurring sporadically under time pressure. Cross-analysis of interview and documentary data revealed meaningful, participant-specific discrepancies between reported and observed practice, particularly with respect to the degree of independent rewriting. These findings contribute to the growing literature on human-AI interaction in academic writing by showing that direct documentary evidence can complement, and at times challenge, self-reported accounts of AI-assisted writing, with implications for institutional AI pedagogy and policy.]]>
Copyrights © 2026
