Recent AI feedback research is shifting the practical question for schools. The issue is no longer only whether a chatbot can produce comments that look helpful. The harder question is whether students can judge those comments, challenge them, and turn them into a better next attempt.
A June 2026 open-access study in Computers and Education: Artificial Intelligence compared AI-generated and human-written personalized formative feedback using 979 feedback responses, then asked 472 STEM students to evaluate AI and human feedback. The researchers found that AI and human feedback were broadly comparable in pedagogical quality, but both often lacked metacognitive elements. Just as important, students’ ratings were shaped more by the perceived credibility of the source than by the quality of the feedback itself.
That finding fits a wider 2026 conversation. In March, a multi-author research meeting report, The Future of Feedback, argued that generative AI could make feedback more immediate and scalable, while leaving unresolved questions about quality, learner agency, teacher roles, and how feedback actually changes student action. In June, Teacher Magazine’s report on AI as a tool for assessment feedback highlighted the same concern in school practice: feedback should improve the learner, not merely polish the submitted work.
Meanwhile, students are already using AI at scale. RAND’s March 2026 report, More Students Use AI for Homework, and More Believe It Harms Critical Thinking, found that the share of middle school, high school, and college students reporting AI help with homework rose from 48 percent in May 2025 to 62 percent in December 2025. The same report found that 67 percent of students agreed that more AI use for schoolwork would harm critical thinking skills.
What Happened
The strongest recent signal is that AI feedback can be useful, but it is not self-sufficient. The ScienceDirect study suggests that well-structured AI feedback may often resemble human feedback in surface quality. That is encouraging for large classes, tutoring programs, and independent learners who need faster responses than a teacher can always provide.
But the same study also points to a weakness that matters for learning. Feedback that explains what is wrong without helping the learner monitor their own understanding can leave students dependent on the next outside judgment. If a learner cannot tell whether a comment is accurate, relevant, too vague, too confident, or aimed at the wrong goal, the feedback loop becomes fragile.
The RAND survey adds urgency. Many students are not waiting for perfect school policies before using AI. They are already asking for help, answers, rewrites, explanations, and corrections. If students do not have a routine for evaluating AI feedback, the tool can become either an authority they accept too quickly or a shortcut they use to avoid thinking.
Why It Matters
Feedback improves learning only when it changes what the learner does next. A comment such as “add more evidence” is not useful until the student can identify which claim needs evidence, what kind of evidence would count, and how to revise without losing the main argument. A math hint is not useful until the learner knows whether it reveals a concept gap, a calculation slip, or a poor strategy choice.
AI makes this problem larger because it can produce fluent feedback instantly. Fluent language can feel credible even when it is generic, incomplete, misaligned with the rubric, or based on a misunderstanding of the student’s work. That does not mean learners should avoid AI feedback. It means they need evaluative judgment: the ability to inspect feedback instead of merely receiving it.
This is also where teacher guidance still matters. Teachers can model how to ask for narrower feedback, how to compare comments against criteria, how to reject an unhelpful suggestion, and how to decide on one revision action. Without that modeling, students may collect more feedback while learning less from it.
The Practical Learning Conclusion
The best use of AI feedback is not “ask once and obey.” It is a short judgment cycle that keeps the learner responsible for the next move.
- State the target before asking. The learner should name the skill first: clearer thesis, stronger evidence, fewer algebra errors, better vocabulary use, or more accurate explanation.
- Ask for diagnosis, not rewriting. A better prompt is “Identify the two highest-impact problems and explain why they matter” rather than “Fix this.”
- Check the feedback against criteria. Students should compare the comment with the rubric, assignment prompt, worked example, teacher instruction, or textbook method.
- Ask one follow-up question. Useful follow-ups include “What evidence supports that suggestion?”, “Which part of my answer led you to that conclusion?”, and “What is one alternative revision?”
- Choose one next practice step. Feedback should end with an action small enough to do immediately: rewrite one paragraph, solve three similar problems, explain one concept aloud, or make one comparison table.
- Keep ownership visible. Students should mark which suggestions they used, which they rejected, and why. This turns feedback from an answer service into a self-regulation exercise.
For teachers, tutors, parents, and adult learners, the practical rule is simple: AI feedback should create better questions before it creates a better product. If the learner finishes with clearer judgment, a sharper follow-up question, and a specific next practice step, the tool supported learning. If the learner only receives a cleaner answer, the feedback may have improved the work while leaving the learner unchanged.