Stance and Feedback Focus in English Writing: A Corpus-Assisted Analysis of Human and GPT-4-Augmented Comments

Generative AI can reshape English writing feedback, yet how human-informed augmentation changes its wording remains insufficiently understood, and textual change should not be mistaken for pedagogical improvement. This study examines how GPT-4 augmentation recontextualizes stance and metalinguistic focus, providing evidence for critically evaluating AI-supported feedback. A retrospective corpus-assisted analysis compared 4,917 complete human–augmented feedback pairs from the second LEAF release; the human comments came from a single educational consultant. Methods combined paired lexical profiles, markup, quotation and deduplication sensitivity checks, length-aware lexical diversity, an exploratory two-model audit of 64 pairs, and four source-linked close readings. Under the original tokenization, augmented comments averaged 165.2 tokens longer; hedging-associated and positive-evaluation densities increased in 82.0% and 79.1% of pairs, respectively. Task-reference density fell in 67.5% of pairs despite greater task-lexicon presence, while grammar/form and argument/meaning densities increased. Conventional type–token ratios fell, but MATTR-50 increased. Contextual analyses indicated that criticism and task-directed recommendations persisted. The findings describe expansion and reformulation, not demonstrated feedback quality, learner agency or learning gains; their generalizability is limited by the single-contributor, human-conditioned pipeline.