Teacher feedback vs. AI-assisted peer feedback in L2 writing: A quasi-experimental study in a Chinese university

Synthesis: Grounded in Vygotsky's sociocultural theory, this 8-week quasi-experiment treated teacher feedback and AI-assisted peer feedback as two kinds of scaffold and compared them within one Writing course at an application-oriented Chinese university. The control class (n = 31) received teacher written annotations; the experimental class (n = 30) worked in pairs, using DeepSeek to check a partner's draft against the same five-dimensional checklist before negotiating revisions. Across 244 graded texts the two modes behaved as "heterogeneous scaffolds": teacher Feedback offered broader coverage and adapted its focus, while AI-assisted peer feedback held a narrower but steadier focus and produced a stable learning-curve in scores — after trailing in the first task it passed the teacher-feedback class in second-task average revision score. Neither mode improved syntactic complexity, and the AI-assisted advantage in lexical diversity faded after the first task. The authors argue for a hybrid "AI-Peer-Teacher" model rather than substitution.

Key Findings

1. 61 second-year English majors (30 experimental, 31 control) wrote 244 texts; baselines were comparable (82.68, SD = 5.48 vs. 81.56, SD = 5.65; t = 0.79, p = 0.415).
2. First-task feedback: 316 teacher items vs. 185 for AI-assisted peer feedback, the gap in vocabulary errors (t(59) = −2.677, P = .010, d = 0.66) and technical details (t(59) = −2.620, P = .012, d = 0.58).
3. By task 2 teacher feedback fell to 169 items (organization and content/argumentation = 46.15%); AI-assisted peer feedback produced 120 and cut content suggestions (teacher M = 1.11, SD = 1.21 vs. peer M = 0.40, SD = 0.58; t(59) = −2.68, P = .010).
4. Teacher feedback led after task 1 (4.06 vs. 3.87 points) but fell 32.8% to 2.73 points in task 2; AI-assisted peer feedback held at 3.89 points, its task-2 revision mean (89.82, SD = 5.45) exceeding the control (88.35, SD = 5.52).
5. AI-assisted peer feedback raised lexical diversity in task 1 (F(1,59) = 12.288, p < .001, ηp² = .200) but not task 2 (F(1,59) = 1.084, p = .303, ηp² = .022).
6. Syntactic complexity moved under neither mode: only a task-1 time effect was significant (F = 6.184, p = .016; composite mean 2.03 → 2.14); task-2 time, group, and interaction effects were non-significant (F(1,59) = 0.234, p = .630; F(1,59) = 1.991, p = .165; F(1,59) = 0.417, p = .522).
7. The paper's output is an "AI-Peer-Teacher" model: AI for drafting-time content guidance and error marking, peers for revision negotiation, teachers for complex syntax and argumentation.

Study design: one varied factor, two scaffolds

The study ran 8 weeks with 61 second-year English majors at a Chinese application-oriented university. Two classes matched on pretest scores became the experimental group (n = 30, AI-assisted peer feedback) and the control (n = 31, teacher feedback), with gender split 4/26 vs. 4/27 and all other teaching variables constant; only 12 of 61 had used ChatGPT before.

Both classes wrote two ~200-word essays on iWrite 2.0, then revised. The control class received teacher annotations against a standardized five-dimension feedback checklist covering structure, content, grammar, vocabulary, and technical details; experimental pairs used the Generative AI tool DeepSeek on the same checklist and negotiated revisions, with no teacher content feedback. Content validity was 0.92 with three experts (tenure above 10 years) and pilot reliability Cronbach's α 0.78–0.86 on 10 non-experimental essays; Coh-Metrix 3.0 extracted features, SPSS 27.0 ran statistics.

What each mode attended to — and how it shifted

The authors call the two modes "heterogeneous scaffolds": the teacher pattern an "expert dynamic scaffold" re-prioritizing against the learner's zone of proximal development, the AI-assisted one "core-focused" and steadily guided — strong on content and structure, weaker on language form. Unlike static automated writing evaluation, the tool is dialogic and multi-round, not a fixed score.

Divergent trajectories in writing scores

The paper reads authoritative error correction as a fast early gain that may encourage passive, error-avoidance revision, and links error emphasis without praise to reduced motivation and risk-taking; the AI-plus-peer loop has students evaluate suggestions and internalize the tool as a cognitive resource — a Student-AI Interaction path of scaffolding that fades with competence. Gains used repeated-measures ANOVA and within-group t tests: the task-2 between-group difference was small and not significant (mean difference 1.16, 95% CI [−1.06, 3.38], t(59) = 1.048, p = .294); so was the task-1 difference (mean difference −0.19, 95% CI [−3.10, 2.72], t(59) = −0.131); the control class's draft mean was 85.62 (SD = 4.80).

Micro-linguistic outcomes and the hybrid proposal

AI supplies an immediate "resource scaffold" of lexical choices students did not convert into a stable strategy; on draft 1 the AI-assisted class led on LDTTRc (0.80, SD = 0.07 vs. 0.71, SD = 0.06; t = −4.854, p < .001). Syntactic development needs targeted intervention, and language models have not systematically identified learners' syntactic weaknesses. The proposal assigns AI to drafting-time content guidance and error marking, peers to revision negotiation, and teachers to what lies beyond both: complex syntax and argumentation.

What this means for practice

• Course teams. Treat the modes as complementary: AI-assisted peer review holds the line on content and structure while teachers diagnose higher-order problems.
• Instructors. Front-load process training before an AI-assisted peer cycle and add verification prompts for language form.
• Instructional designers. General writing feedback will not move syntactic complexity; design targeted sentence-pattern work over a horizon longer than 8 weeks.
• Administrators. The teacher's adaptive feedback was time-expensive, its early advantage flattening once tasks demanded structural breakthroughs; offload routine error marking to AI.
• Assessment designers. Track holistic scores and micro-linguistic indicators separately; AI-assisted peer review needs intermediary capabilities, so students must challenge tool output rather than let reliance impair Critical Thinking.

Limitations

• The intervention lasted only 8 weeks and drew English majors from a single university, restricting generalizability.
• The study traced feedback patterns and scores, not how learners internalize feedback types, so micro-level mechanisms stay unexplained; it calls for interviews and think-aloud methods.
• DeepSeek is named without a version identifier, and consent ran 14 March to 27 March 2025; results describe a superseded model generation — a 2025 snapshot.
• Intact classes rather than random assignment, and the second-task between-group difference did not reach significance; score findings rest on within-group change.

Citation

Tang, J., Li, P., & Luo, R. (2026). Teacher feedback vs. AI-assisted peer feedback in L2 writing: A quasi-experimental study in a Chinese university. PLOS ONE.