Essay-writing proficiency is an important indicator of English language competence among English as a Foreign Language (EFL) learners. However, providing written corrective feedback manually can place considerable demands on instructors’ time and workload, particularly in large classes. This study evaluates the performance of a Generative Pre-trained Transformer (GPT) model in detecting and correcting grammatical and lexical errors in EFL students’ argumentative essays, while also exploring how students perceive and respond to the feedback it provides. An exploratory descriptive qualitative design was employed with 15 English Language Education students from three semester levels (1, 3, and 5). Essay drafts and semi-structured interview transcripts were analyzed using the Error Analysis framework (Corder, 1981; Ellis, 2008) and Thematic Analysis (Braun & Clarke, 2006). The findings show that GPT achieved its highest correction accuracy in Semester 1 students’ essays (78%), whereas the rate of over-correction increased to 20% in Semester 5 as sentence structures became more complex. The qualitative findings reveal a clear difference in students’ metacognitive responses. Semester 1 students tended to exhibit automation bias by accepting GPT-generated corrections with little critical evaluation, while Semester 3 and 5 students demonstrated greater learner agency by evaluating, filtering, and negotiating the suggested corrections. Overall, the findings suggest that GPT can serve as a useful form of scaffolding for formative feedback, but its integration into EFL writing pedagogy should be accompanied by critical AI literacy to help students maintain their metalinguistic awareness and authorial voice