From errors to excellence: AI-driven scaffolding for advancing L2 writing skills

This study leveraged large language models’ (LLMs) function of mimicking students’ particular errors and generating essays that embed those errors, thereby yielding targeted revision exercises. This research examined how AI-generated feedback and the subsequent error correction exercises influenced undergraduates’ academic writing in an eight-week academic writing course in mainland China, also probing the mediating role of proficiency levels. A convenience sample of 44 English-majors was allocated to an experimental group (n = 25) that received dual interventions tailored to each learner’s recurrent writing problems, or a comparison group (n = 19) that received AI-generated feedback only. Subsequently, a purposive sub-sample of 12 participants was selected for online semi-structured interviews for qualitative insights. The quantitative analysis revealed significant improvements in both groups’ writing after the intervention, with proficiency level moderating the outcomes. Crucially, the experimental group significantly outperformed their comparison counterparts in mid-test. By deploying GPT-4o to simulate students’ recurrent writing errors and generate targeted correction materials, this study informs pedagogy on how LLMs scaffold personalized learning environments, enabling differentiated instruction tailored to the proficiency continuum in academic writing courses.