What happened
Recent research on AI writing feedback is moving the discussion away from a simple question: can an AI tool comment on a draft? The more useful question is now: does the feedback help students do the thinking that revision requires?
One important 2026 study from University of Michigan researchers tested an AI-mediated feedback system called FeedbackWriter in a large undergraduate economics course. In the randomized trial, teaching assistants received AI-generated, rubric-aligned suggestions while they reviewed students’ essays, but the human graders kept control over what to accept, edit, or reject. The study reported that students who received AI-mediated feedback produced higher-quality revisions than students who received human-only feedback, with stronger gains when teaching assistants adopted more AI suggestions. Read the university summary here: AI helps instructors give better feedback but can’t replace them. The paper is also available here: AI-Mediated Feedback Improves Student Revisions.
A second 2026 study in the International Journal of Educational Technology in Higher Education adds a caution. Researchers compared teacher feedback with GenAI feedback generated through different prompting methods for argumentative essays. Chain-of-thought prompting produced structurally strong feedback, but that did not automatically produce better revised essays. Students improved across feedback conditions, and the authors concluded that feedback quality alone is not enough; students’ engagement with the feedback and uptake during revision are critical. Read the study here: Generative AI offers more, but students revise less.
Another 2026 article in Frontiers in Education examined how pre-service teachers viewed AI-assisted writing tools for student self-assessment. The study found cautious and mixed attitudes. Teachers with direct experience using the tools were more favorable, especially for specific skills such as spelling and composition, but concerns remained about critical thinking and over-reliance. Read it here: Use of AI-assisted writing tools for student self-assessment at school.
There is also a growing equity warning. Education Week recently reported on Stanford research showing that AI writing feedback could change when models were given demographic or motivational descriptions of the student, even when the writing sample itself was the same. The report described differences in tone, emphasis, and level of critique across race, gender, and motivation labels. Read the report here: AI Changes Its Feedback on Students’ Writing When It Knows Their Race, Gender.
Why it matters
The common thread is that AI feedback is becoming more capable, but revision is still a learner activity. A tool can notice missing evidence, suggest clearer structure, flag weak transitions, or point to rubric criteria. None of that guarantees that the student has planned the argument, retrieved what they know, checked the source, or rewritten the idea independently.
This distinction matters because revision is not proofreading. Real revision asks a learner to decide what the text is trying to do, compare the draft with a goal, diagnose the gap, and make a better choice. Those steps involve metacognition, retrieval, source evaluation, and transfer. If students simply accept AI edits, the final paragraph may improve while the underlying revision skill remains underdeveloped.
The Michigan study is encouraging because AI was used behind the scenes to support human graders, not to replace them. The system gave teaching assistants more structured options, while teachers still judged whether the feedback fit the assignment and the student draft. That is a different learning design from handing every student a chatbot and assuming the comments will be accurate, fair, and instructionally useful.
The Springer study explains why that distinction is important. Strong-looking feedback does not automatically become strong revision. A comment can be detailed and still fail to change how a student thinks about the draft. The learner has to interpret the comment, decide whether it is right, connect it to the assignment goal, and rewrite in a way that preserves their own meaning.
The bias concern makes source checking and teacher judgment even more important. If AI feedback shifts because irrelevant student labels are present, then students and educators need tighter rules about what information goes into prompts. A writing tool should respond to the writing task, rubric, evidence, and draft, not to assumptions about who the student is.
The learning conclusion
The practical conclusion is not to avoid AI writing tools. It is to use them in a way that keeps the student responsible for the thinking work of revision.
A good AI feedback routine should protect four skills:
- Planning. Before asking for feedback, the learner should write the purpose, audience, claim, and main evidence in their own words.
- Retrieval. Before reading AI suggestions, the learner should list what they already think needs improvement without looking at the tool’s answer.
- Source checking. Any AI suggestion about facts, examples, quotations, or evidence should be checked against the original source or assignment material.
- Independent rewriting. The final revision should be written by the learner, not pasted from the tool, with at least one sentence explaining why the change improves the draft.
Students can use a simple three-pass method. First, do a self-review: mark one unclear claim, one place that needs evidence, and one sentence that sounds weak. Second, ask the AI for feedback tied to the rubric, but request questions and priorities rather than a rewritten essay. Third, choose two changes, check them against the assignment and sources, and rewrite the paragraph independently.
Teachers can make this routine visible by asking students to submit a short revision note with the final draft. The note can include the original sentence, the feedback they considered, the decision they made, and the reason for the change. This turns AI feedback from a hidden shortcut into an object for reflection.
Parents and tutors can ask better questions too. Instead of asking, “Did the AI fix it?” ask, “What did you decide to change, and why?” That question brings the learner back to judgment. It also reveals whether the student understands the writing problem or is only following automated instructions.
The strongest use of AI writing feedback is as a second reader, not a substitute writer. Let the tool surface possibilities, but make the learner plan, retrieve, verify, and rewrite. That is how AI feedback can support revision skills without quietly replacing them.