The cumulative effects of AI-generated feedback on syntactic development in high school EFL writers: a longitudinal quasi-experimental study

Abstract

Introduction While AI-generated writing feedback research has flourished, most studies prioritizeholistic quality over syntactic competence and rely on static pre-post comparisons, neglecting dynamic multi-round evolution. Methods This 35-week quasi-experiment examined cumulative associations between exposure to supplementary AI-generated feedback components and syntactic development among 69 Chinese high school EFL learners (Group A: n = 33 receiving automated AI scoring, official model compositions, and supplementary AI-generated feedback components; Group B: n = 36 receiving automated AI scoring and official model compositions). Piecewise regression, cognitive load scales, and qualitative triangulation were used across 10 timed continuation tasks. Results The study described non-linear trajectories: an initial adaptation decline followed by faster gains. Among the 10 target error types, only Part-of-Speech errors showed a statistically significant reduction after FDR correction; other error categories exhibited only suggestive trends that did not remain robust after multiple comparison adjustment. Mixed-methods analysis described a hypothesized cognitive load pattern in which extraneous load diminished yet germane processing diverged contingent on learner strategy, with passive offloading coinciding with an illusion of competence. Four revision strategies emerged. Discussion The findings are consistent with a conditional pathway within the CLT-SAT framework as an interpretive lens and suggest the potential of a phased Human-in-the-Loop model for contexts where supplementary AI-generated feedback is provided, one that scaffolds autonomy while guarding against passive offloading.