Peer feedback leads, GenAI complements: tracing revision patterns in undergraduate projects

Generative artificial intelligence (GenAI) is increasingly used to provide formative feedback in higher education, yet less is known about how students translate AI-generated suggestions into enacted revisions, particularly compared with peer feedback. This mixed-methods study examined how 184 undergraduate students working in small groups revised course projects after receiving formative feedback from peers and ChatGPT, and whether revision behavior varied according to initial project quality. Final submissions included revision annotations documenting implemented changes, attributed source(s), and rationales. An analysis of 656 annotated revisions revealed four key findings. First, peer feedback drove the largest share of attributed revisions, followed by ChatGPT and combined peer-ChatGPT sources. Second, content analysis identified nine revision types, with conceptual precision and content elaboration being the most prevalent. Third, revision type and feedback source were significantly associated, a pattern we term source-revision alignment. Fourth, initial project quality was related to source reliance rather than revision quantity, a contrast we term agency-dependency divergence. Groups with high-quality initial submissions relied significantly less on ChatGPT and engaged in more self-initiated revisions. These findings highlight the complementary roles of peer and ChatGPT feedback in supporting revision and underscore the importance of considering initial performance when integrating GenAI into formative feedback designs.