A prompted path to training ChatGPT for targeted feedback on B1-level paragraph writing in relation to student uptake

2025-9-19
Yıldırımer, Gülçin
This study aimed to explore the general performance of trained ChatGPT in providing feedback on B1-level EFL students' paragraph writing and examine students' uptake patterns of AI-generated feedback. ChatGPT 4o was systematically trained through a five-phase prompt engineering process to provide structured feedback on both local and global errors in student paragraphs. Data were collected from 24 participants through student paragraphs, ChatGPT-generated feedback, and think-aloud protocols. The study examined (1) the overall performance of trained ChatGPT in identifying and addressing both local (language use and mechanics) and global (content and organization) writing issues, (2) the extent to which learners utilized this feedback in their revisions, alongside the reasons for unattempted revisions. The findings revealed that the trained ChatGPT consistently provided criteria-based, structured feedback addressing both local and global writing aspects with an appropriate supportive tone for B1 learners. The model also demonstrated high accuracy in feedback provision; however, there were also instances of inaccurate or unidentified feedback. Moreover, students engaged most with feedback provided for local issues, while uptake of global feedback was comparatively limited. In addition, think-aloud protocols identified two primary reasons for unattempted revisions: comprehension difficulties and disagreement with ChatGPT’s assessment. This research aims to contribute to understanding AI-assisted writing feedback and provide practical insights into integrating AI feedback into L2 writing pedagogy.
Citation Formats
G. Yıldırımer, “A prompted path to training ChatGPT for targeted feedback on B1-level paragraph writing in relation to student uptake,” M.A. - Master of Arts, Middle East Technical University, 2025.