ABSTRACT The integration of artificial intelligence (AI) into second language writing pedagogy necessitates a rigorous comparison of its efficacy against traditional teacher feedback. This quasi‐experimental study investigated the differential effects of AI‐only (ChatGPT‐4‐generated feedback), teacher‐only, and no‐feedback conditions on the longitudinal development of lexical richness in the writing of 226 Chinese English majors. Over a 12‐week intervention, three longitudinal learner corpora were compiled from participants' pretest, immediate posttest, and delayed posttest essays. Computational analyses using mixed‐effects modeling (linear mixed‐effects model and generalized linear mixed‐effects Model) tracked development in lexical diversity (moving‐average type–token ratio), density, and sophistication (Lexical Frequency Profile). Results revealed a striking divergence in developmental patterns. The AI‐only group demonstrated significant, robust, and sustained growth across all three dimensions, breaking the “lexical ceiling” often observed in L2 writing. Notably, AI feedback prompted a near‐doubling of lexical sophistication probability that was maintained 4 weeks later. In contrast, teacher feedback produced significant but transient gains in density and sophistication, which regressed to baseline levels at the delayed posttest, though it successfully sustained lexical diversity. The control group showed minimal improvement. The findings underscore the significant potential of AI‐generated feedback to provide the consistent, data‐driven scaffolding necessary for profound and lasting lexical development, suggesting a re‐evaluation of feedback dynamics in the L2 writing classroom. This study contributes to the fields of SLA and educational technology by offering a nuanced, corpus‐based longitudinal perspective on automated writing evaluation.
Xie et al. (Tue,) studied this question.