Lighting the spark: How LLMs empower underprivileged Chinese high school students to reimagine EFL writing

This study examined the impact of introducing DeepSeek, a locally accessible large language model, to supplement insufficient writing feedback in an under-resourced Chinese high school. Using a quasi-experimental pretest-posttest design, two intact Grade 11 classes were randomly assigned to receive DeepSeek-generated feedback (n = 22) or conventional teacher feedback (n = 24) over nine weeks. Beyond writing scores, we traced changes at three levels: behavioral (feedback uptake), psychological (writing self-efficacy), and metacognitive (self-regulated learning strategies). Quantitative results showed that after LLM feedback was introduced, students successfully implemented more revisions and made meaningful gains in writing performance, self-efficacy, and self-regulation. However, qualitative findings revealed that while the resource gap was narrowed, it was far from fully closed. The LLM tool can compensate for the basic shortage of feedback, yet its supplementation is inherently skewed toward surface-level language issues and low-inference suggestions, contributing little to higher-order concerns like argumentation and idea development. Students in under-resourced settings do not lack potential but the conditions to realize it, and these conditions alone do not guarantee better learning. Narrowing the gap between existing and ideal instructional environments requires not only introducing the tool but guiding its use. Deeper, sustainable, tool-independent competence demands higher-order guidance beyond the tool itself.