Teacher feedback VS AI-assisted peer feedback in L2 writing: A quasi-experimental study in a Chinese University

A peer who discusses an essay with AI support is doing something different from a chatbot that comments directly on a student's work. A PLOS One study examines that combined activity across two university English classes. The results suggest that teacher feedback and AI-assisted peer feedback can produce different revision patterns, while also showing how easily broad claims about language development can outrun the measures collected.

Jinfeng Tang, Ping Li, and Rong Luo followed 61 second-year English majors at a university in central China for eight weeks. Two intact classes were selected from six using initial writing performance: thirty students received AI-assisted peer feedback and thirty-one received teacher feedback. The same instructor, teaching content, and lesson periods were maintained. This was a quasi-experiment, however, with no individual random assignment. A similar starting score does not remove every possible difference between classes.

Each student completed two argumentative writing tasks, with an initial and revised draft for each, yielding 244 texts. Both feedback arrangements used a common checklist covering organization, argumentation, grammar, vocabulary, and technical details. In the experimental class, students worked in pairs and used DeepSeek to analyze a partner's draft, then discussed and negotiated the comments they would offer. The teacher supervised that process without directly supplying content feedback. The comparison class received the instructor's written comments and revision suggestions.

Teacher feedback covered more issues in the first task, with 316 recorded suggestions compared with 185 in the AI-peer class. Its focus changed in the second task toward organization and argumentation. Essay scores improved under both approaches, but the reported revision gains followed different trajectories. The AI-peer class's gains were relatively stable across the two tasks, whereas the teacher-feedback class showed a larger initial gain and a smaller subsequent one. These are results for successive drafting and revision cycles, rather than evidence from a delayed independent transfer test.

The language-feature analyses were more qualified than a general statement that AI improved writing. The AI-peer class showed an initial lexical-diversity advantage, but the group difference was no longer significant in the second task. For syntactic complexity, the detailed results reported an overall improvement across the first revision cycle, without a significant group difference or interaction. In the second task, the reported time, group, and interaction effects were all non-significant. There was therefore no established advantage of one feedback mode for sentence complexity.

For teachers, the interesting design element is the discussion between classmates after consulting AI. Learners must decide which suggestions fit a partner's meaning and explain those choices. A classroom adaptation could retain a common rubric and ask students to document why they accepted or rejected a suggestion. That is an instructional implication, not a mechanism demonstrated by this study: it did not closely trace how learners internalized feedback or separate the effect of AI from peer negotiation.

The authors propose combining teacher, peer, and AI contributions. That hybrid was not itself tested here. The evidence comes from one institution, two non-randomized classes, two writing tasks, and platform-based writing scores. It supports further investigation of complementary feedback roles, while leaving durable writing development, independent performance, and the specific value added by the AI component unresolved.