AI and peer feedback can work together in a structured writing course. In Wen Liu’s (2026) study of university students learning English in China, students who revised with AI before exchanging feedback with classmates showed greater writing gains and engagement than students using peer feedback alone. The results support examining this combined approach, although the additional feedback and revision time makes AI’s individual contribution difficult to establish.
What deserves attention is the instructional sequence. Students began with their own drafts, received guided AI comments, revised, and then discussed their writing with peers. Their engagement developed within a course that included training, assessment criteria, and instructor oversight.
Giving students feedback they can use
Peer feedback asks students to read another person’s writing and explain what works, what is unclear, and what could improve. In a second language classroom, that can be demanding. Learners may notice a problem without knowing how to describe it, or hesitate to question a classmate’s suggestion.
AI introduces another source of comments, often available immediately. Yet receiving more feedback does not tell us whether students understand it or can use it thoughtfully. A learner might replace a sentence without grasping the reason for the change.
The question Liu addresses is whether combining AI feedback with peer interaction can support both engagement and writing performance. This connects with student feedback literacy: the capacities students need to interpret comments, make judgements, and act on feedback. The design of the activity matters because it determines what learners do with the advice they receive.
What the AI and peer feedback study found
Liu studied 122 English-as-a-foreign-language students from three existing classes at one university in southern China. Two classes formed the combined-feedback group, while the third used conventional peer feedback. Over a ten-week intervention within a twelve-week study, students completed five writing assignments.
In the combined group, students submitted an initial draft to ChatGPT using prompts tied to four assessment areas, revised, and then participated in individual peer review and small-group discussion before revising again. The instructor provided preparatory training and periodically reviewed students’ AI interactions.
The combined group showed greater improvement in task achievement, coherence and cohesion, vocabulary, and grammatical range and accuracy. Engagement questionnaires also favoured this group across behavioural, emotional, and cognitive dimensions.
Interviews, reflective journals, and analysis of drafts added detail: some students described greater willingness to revise, checked suggestions against other resources, and made a wider range of changes. Others reported difficulty understanding AI comments or receiving useful advice on organisation. These accounts help explain students’ experiences; they do not establish that every learner evaluated feedback critically.
The design limits what we can conclude. Students were allocated through existing classes, and all classes came from one institution. The combined group had an additional AI feedback and revision stage, so the comparison does not isolate AI from extra practice or time.
There was no delayed test to establish whether improvements lasted, and the author acknowledges possible novelty effects. The evidence concerns a supported classroom process, with writing gains measured during the study.
Designing revision around student decisions
I would begin with the sequence used in the study. Students wrote first, consulted AI, and then brought their revised work into peer discussion. This creates several opportunities to return to the developing argument and consider how a reader understands it. It also gives teachers a way to define AI’s role before students use it.
The researchers recommend training students and educators to evaluate AI feedback critically. In practice, I would model that evaluation with a short sample paragraph. A class could examine whether a suggestion addresses the assessment criteria, preserves the intended meaning, and supplies an explanation the writer understands. This is my suggested adaptation, not an activity tested separately in Liu’s study.
A manageable revision record could help. Ask students to select one consequential suggestion and explain what they changed and why. They could also record a suggestion they rejected.
This would give us evidence of their reasoning without requiring an account of every edit. My discussion of feedback literacy in AI-enhanced writing considers related questions about how students engage with assistance during composition.
Peer discussion needs a clear purpose too. After an AI-supported revision, classmates could examine whether the central claim is convincing, whether an example supports it, or where the reader needs more context.
Teachers would still need to prepare students to give specific comments. We should avoid assuming that AI will reliably handle language while peers automatically provide strong advice about ideas.
For second language learners, preserving intended meaning deserves particular care. A suggested sentence may sound fluent while changing the strength of a claim or removing something the writer wanted to express. Asking the student to explain the change, and inviting a peer to describe how they read it, could make that shift visible.
Finally, I would include a later writing task with clearly defined assistance conditions. Liu’s study does not establish long-term retention, so a subsequent task could help a teacher examine what students can apply after the feedback cycle. Within AI and teaching practice, this is a useful question to keep asking: what can learners explain and use after receiving support?
Reference
Liu, W. (2026). AI-peer integrated feedback in second language writing classes: Exploring students’ engagement and writing performance. Frontiers in Psychology, 17, Article 1871951. https://doi.org/10.3389/fpsyg.2026.1871951
