Mediating L2 writing revision: The impact of teacher-moderated ChatGPT feedback on L2 accuracy and lexical diversity

Studies that empirically investigate the concrete linguistic impact of GenAI on L2 writing remain scarce. This study addresses this gap by examining 91 in-class ESL texts alongside their 91 revised versions, produced using ChatGPT feedback in a teacher-moderated context. We investigate linguistic changes in L2 accuracy, namely learners’ ability to produce error-free output, and lexical diversity, the range of distinct word types used in writing. Learner corpus research methods were used to do so: errors were manually annotated with the Louvain Error Taxonomy (Granger et al., 2022), while lexical diversity was measured automatically with TAALED (Tool for the Automatic Analysis of Lexical Diversity; Kyle et al., 2021). The findings reveal that, although ChatGPT-assisted revisions showed some improvement in certain areas of accuracy, the texts still contained a substantial number of errors. Lexical diversity increased more visibly. Results are analysed against the structured pedagogical safeguards that guided learners’ interactions with ChatGPT. To avoid wholesale cutting and pasting from the tool, the students (i) co-developed and signed an ethical use agreement; (ii) were provided with teacher-suggested sample prompts which they extensively relied on and (iii) were required to submit ChatGPT chatlog interactions to the instructor. Together, these safeguards encouraged focused revisions rather than maximal rewriting, shaping the nature and extent of the changes observed in the revised texts. Drawing on our results, the study critically reflects on the strengths and limitations of the pedagogical intervention, showing how controlled AI use can address concerns about authorship, L2 assessment and ethics.