Where revision actually happens

A short report landed on 14 October that I had to sit with, because it stated plainly a question I had been circling for a few weeks. Research Today covered a study by Wen Liu, published 21 September in Frontiers in Psychology, of 122 Chinese learners of English writing across a 10-week intervention. One group revised essays with peer feedback alone. The other revised with AI feedback and peer feedback together. The combined group's writing improved more across the four measured dimensions, and their engagement with the revision ran higher (Liu, Frontiers in Psychology, 21 September 2026; reported by Research Today, 14 October 2026).

The headline reads like a win for AI feedback. The finding sits one level lower than that, and it is a finding this publication has been pushing toward for a few weeks.

Across three studies this year that compared AI feedback with teacher feedback on real student writing, the pattern is the same. When researchers raise the quality of the feedback and change nothing else, revisions do not improve. When researchers change what the student has to do with the feedback, revisions improve regardless of its source.

The pen revises. The feedback can only point.

Three studies, one shape

Farrokhnia and colleagues ran a randomized three-group experiment at Wageningen University with 70 graduate students writing argumentative essays. AI feedback produced with chain-of-thought prompting was rated higher in quality than the teacher's. The students who received it did not revise their essays any better than the students who received the lower-rated teacher feedback (International Journal of Educational Technology in Higher Education, Vol. 23, Issue 1, December 2026; return from edition 12). Higher-quality input, flat output.

Lu, Ju, Dudley, Sano and Wang ran a randomized trial with 354 undergraduates in a large introductory economics course at the University of Michigan, with eleven teaching assistants grading 1,366 essays. Half the essays were given human-only feedback. The other half were given AI-mediated feedback: TAs received model suggestions and chose which to adopt or edit before writing their own comment. Students who received AI-mediated feedback produced significantly higher-quality revisions, with a Cohen's d of 0.50, and the effect grew as TAs adopted more AI suggestions (Lu et al., arXiv:2602.16820, 18 February 2026; conference version at DOI 10.1145/3772318.3791121). The AI changed the first draft of the feedback. The student still received a human's judgment.

Alsaiari and colleagues ran a quasi-experimental study across three cohorts, 13,037 students and 51,296 submitted resources. Students in a Directed Feedback condition received AI feedback without structured support. Students in a Self-Directed condition could opt into AI dialogue. Students in an Enacted Feedback condition were prompted to select feedback, evaluate it, and respond. The modelled probability of feedback uptake ran at 14.1 percent in Directed, 0.1 percent in Self-Directed, and 26.2 percent in Enacted (Alsaiari et al., arXiv:2608.11625, 12 August 2026). Same model, same subject matter. The thing that moved uptake was what the student had to do with the feedback.

Caveats, up front. The three studies run on different populations and different fields, with different measures of "revision quality". The Lu study is a conference paper and I read the arXiv version. The Alsaiari figures are modelled probabilities from a quasi-experimental design with self-selection into cohorts. The Liu study compares two configurations rather than isolating AI's independent contribution. No single study is proof. The shape repeats, which is why I will argue the shape.

Why the quality of the suggestion is not the lever

A piece of feedback is a sentence written about a student's draft. A revision is a sentence written by the student, who has to decide what to change, make the change, and sometimes rewrite the paragraph three sentences earlier that the first change broke. Only one of those activities is where learning gets built, and the student is doing it.

The quality of the suggestion changes what is theoretically available to the reviser. It does not change whether the reviser is doing the revising. A student who copies a suggestion into their draft without reading the paragraph it lands in has not revised. A student who reads a mediocre teacher comment, pokes at the paragraph for ten minutes and finds a stronger way to say what they meant has.

What the three studies suggest a workflow actually needs

Three design levers that moved revision in the studies above. Mine, drawn from reading them together and untested as a set. Edit them to fit the setting.

• Make the student choose which pieces of feedback to act on, and say why. Alsaiari's Enacted condition did this and shifted uptake from near zero to roughly one in four. A rubric for select, justify, apply takes less than a page.
• Put a human between the model and the student when the stakes matter. Lu's teaching assistants were not grading faster; they were grading differently. The AI drafted candidate feedback. The TA decided what reached the student. The Cohen's d of 0.50 came out of that arrangement, and it is the largest revision effect across the three studies here.
• Separate the revision act from the reading of the feedback by at least one working session. Flagged, because this one sits outside the three studies above. The older revision literature finds that students who revise within an hour of receiving feedback incorporate it mechanically, and students who revise after a day make structural changes. A workflow that collapses the two steps collapses the second kind of change.

All three ask something different of the student. The tool is the same tool.

A thread, third pass

Edition 12 argued that the order a student brings an AI in predicts their performance on the assignment. Edition 13 argued that a tutor's dashboard counts the moment the question stops rather than the moment learning starts. This edition argues the same thing from a third angle. Better feedback is a floor. The ceiling is set by what the student is asked to do with it.

How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.

One question I would like answered, and I do not have a good one of my own. If you teach, or support people who do, where in your assignment design is the step that asks the student to do something with the feedback rather than just read it? If it is implicit, how often do the students actually notice it is there? The published literature says a lot about feedback and very little about the step after.

Next edition: beginning a short arc on where learning hides inside the activities everyone already does. First subject: the question a student asks in silence.

These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.

Common questions

Does better AI feedback improve how students revise? Not on its own, in the studies reviewed here. In a randomized experiment with 70 graduate students at Wageningen University, AI feedback was rated higher in quality than the teacher's, and the students who received it did not revise their essays any better than the students who received the lower-rated teacher feedback.

What did move revision quality in these studies? What the student was asked to do with the feedback, and whether a human stood between the model and the student. In the Michigan trial, teaching assistants chose which AI suggestions to use before writing their own comment. In the large uptake study, students who were prompted to select, evaluate and respond to feedback acted on it far more often.

How firm are these findings? They come from different populations, fields and measures of revision quality, and the uptake figures are modelled probabilities from a design where students sorted into cohorts themselves. No single study is proof. The shape repeats across all three, which is the argument.

References

• Liu, W. (2026, 21 September). The impact of integrating AI and peer feedback on EFL learners' writing engagement and performance. Frontiers in Psychology, 17. https://doi.org/10.3389/fpsyg.2026.1871951 (Primary comparison for the opening. Metadata and summary read from the Research Today report; the full Frontiers article page was not opened for this edition.)
• Research Today. (2026, 14 October). Study of 122 students finds AI and peer feedback together improved English writing. https://researchtoday.co.za/study-of-122-students-finds-ai-and-peer-feedback-together-improved-english-writing/ (Confirms the 21 September publication date and the 122-student, 10-week study design.)
• Lu, X., Ju, K. P., Dudley, M., Sano, L., & Wang, X. (2026, 18 February). AI-Mediated Feedback Improves Student Revisions: A Randomized Trial with FeedbackWriter in a Large Undergraduate Course. arXiv:2602.16820. https://arxiv.org/abs/2602.16820 (Author list, N=354, 11 TAs, 1,366 essays, and Cohen's d of 0.50 read from the arXiv listing. The full PDF was not opened. Conference version: Learning @ Scale, DOI 10.1145/3772318.3791121.)
• Farrokhnia, M., Latifi, S., Papadopoulos, P. M., Hogenkamp, L., Gijlers, H., Khosravi, H., & Noroozi, O. (2026). Generative AI offers more, but students revise less: comparing the effects of teacher and AI feedback on student essay revisions. International Journal of Educational Technology in Higher Education, 23(1). https://doi.org/10.1186/s41239-026-00579-9 (RETURN. Cited previously in edition 12. Used here for the AI-feedback-rated-higher finding and the "revise less" result.)
• Alsaiari, O., Baghaei, N., Lodge, J. M., Gašević, D., Winstone, N., & Khosravi, H. (2026, 12 August). Making AI-Generated Feedback Matter: A Large-Scale Study of Feedback Workflows and Student Enactment. arXiv:2608.11625. https://arxiv.org/abs/2608.11625 (Metadata and the 14.1 / 0.1 / 26.2 percent modelled probabilities read from the arXiv listing. The full PDF was not opened.)