ChatGPT's Handling of L2 Learners’ Fossilized Errors: A Linguistic Evaluation

Abstract

This study investigates ChatGPT’s capacity to address fossilized grammatical errors in English as a Foreign Language (EFL) learners’ academic writing. Through a mixed-methods design, a controlled corpus of 500 hypothetical sentences containing persistent error types, such as verb tenses, articles, prepositions, and non-idiomatic expressions, was submitted to ChatGPT-4. Quantitative analysis evaluated correction accuracy using standard metrics (precision, recall, F-score), while qualitative content analysis assessed the pedagogical appropriateness and consistency of ChatGPT’s feedback. Results showed high accuracy in correcting rule-based structures (e.g., subject-verb agreement), but significantly lower performance for context-sensitive and fossilized errors. While ChatGPT often provided clear corrections, its feedback frequently lacked explanatory depth, contextual sensitivity, and scaffolding necessary for promoting learner noticing and long-term acquisition. These findings suggest that although ChatGPT can effectively support surface-level proofreading, it cannot fully substitute the role of human instructors in addressing deeply ingrained L2 errors. The study emphasizes the importance of explainable AI, AI literacy, and hybrid instructional models that combine technological efficiency with pedagogical intentionality. It offers implications for educators, curriculum developers, and AI tool designers seeking to integrate language models into second language acquisition contexts.