A Comparative Study Of AI-Generated Corrective Feedback And Teacher-Written Corrective Feedback Among A2-Level EFL Learners In Tertiary Education

Abstract

This study compared teacher-written and AI-generated written corrective feedback on A2-level English-as-a-foreign-language (EFL) paragraphs to examine coverage, overlap, and pedagogical value. Learner texts (N = 39 paragraphs) received feedback from an experienced teacher and ChatGPT-4 (June 2025 release). Feedback was coded at the category level (Grammar, Vocabulary, Spelling & Punctuation, Syntax/word order) as present/absent for each learner × category pair. Paired analyses showed that ChatGPT flagged more feedback categories per paragraph than the teacher, with a significant within-pair effect (Wilcoxon signed-rank, p <.001; d = 1.15). A McNemar exact test on discordant pairs indicated a significant asymmetry favoring ChatGPT (p <.001), demonstrating that the model contributed substantially more unique category notices beyond those offered by the teacher. Category-specific contrasts revealed that the surplus was concentrated in Syntax and Vocabulary, whereas Grammar and Spelling & Punctuation exhibited near-complete overlap between rater types. The results suggest that large-language-model feedback can broaden the lexical–syntactic feedback net without sacrificing surface accuracy, offering efficiency gains while maintaining complementarity with teacher expertise. Pedagogically, the findings support hybrid workflows in which teachers curate or mediate AI suggestions to preserve motivational tone and help learners prioritize revisions. Limitations include focusing on a single proficiency band and binary category coding; future work should track uptake and learning gains and assess cost–benefit profiles of teacher-mediated AI feedback.