AI AND METADISCOURSE: A CORPUS ANALYSIS OF CHATGPT REVISIONS TO EFL STUDENT WRITING

As research into the potential of artificial intelligence in education has grown, several studies have explored the impacts of AI on students' writing. However, crucial aspects of students' academic writing and how AI influences them remain understudied. This gap is unfortunate because the increasing number of students and teachers using AI models for academic writing requires understanding their roles and potential. To address this, the present study examines how ChatGPT uses metadiscourse when revising students' writing. Using a corpus of 240 texts with 85000 tokens written by EFL students and revised by ChatGPT 4o, the study applied a corpus linguistics method to examine the normalized frequency and percentage ranges of Metadiscourse Markers (MDMs) as categorized by Hyland's (2005) framework in both original and revised texts. It also used T-tests to assess the statistical significance of differences in MDM use across the two subcorpora. Results reveal varied distribution patterns across different MDM categories and subcategories. Interestingly, AI revisions tend to emphasize interactional resources more than text-organizing devices. At the subcategory level, attitude markers increased significantly after AI intervention, while transitions significantly decreased in the AI versions. However, all other differences are not statistically significant, indicating that MDM usage in students' writing and AI revisions remains generally comparable. The findings suggest that AI revisions should be complemented by human oversight and can effectively support the development of academic writing. They provide insights into the potential of AI interventions and open avenues for further research in broader settings. It is recommended that future research consider writing from other disciplines, adopt a longitudinal design, and test other common AI models.