On this page
Feedback is one of the most useful things a teacher gives and one of the most time-consuming. That makes it an obvious job to hand partly to AI. The first controlled studies are now in, and they give a more useful answer than either the hype or the fear: AI feedback helps in specific, measurable ways, it is not yet as good as a careful teacher, and it works best when a person stays in charge.
What have controlled studies found so far?
AI feedback versus no feedback
In one experiment, 459 upper secondary students learning English as a foreign language wrote an argumentative essay (Meyer and colleagues, 2024). One group revised using feedback generated by GPT-3.5; the other revised with no feedback. The AI-feedback group improved their essays more (d = 0.19) and reported higher task motivation (d = 0.36) and more positive emotions (d = 0.34). Those are modest effects, and the comparison was with no feedback, not with a teacher.
AI feedback versus human feedback
A research team compared ChatGPT's feedback with human feedback on 200 secondary student essays (Steiss and colleagues, Learning and Instruction, 2024). They scored each on five qualities: being criteria-based, giving clear directions, accuracy, prioritising what matters, and supportive tone. Humans were better on every quality except being criteria-based. The authors judged the practical gap to be not substantial, given how easily the AI feedback was produced.
AI that coaches the human
A large trial kept people at the centre. Tutor CoPilot gave human tutors AI suggestions during live tutoring sessions, in a randomised trial with 900 tutors and 1,800 K-12 students from under-served communities (Wang and colleagues, 2024). Students whose tutors had access were 4 percentage points more likely to master topics, and 9 points more for students of lower-rated tutors. Tutors with the tool were more likely to ask guiding questions and less likely to give away answers. Tutors also reported suggestions that were not right for the grade level.
Saving teacher time
In England, an Education Endowment Foundation trial gave Key Stage 3 science teachers ChatGPT plus a short guide for lesson and resource preparation (EEF report, December 2024, summarised by Stanford SCALE). They spent 56.2 minutes a week on planning against 81.5 minutes for the comparison group, a 31% reduction, and an expert panel found that quality did not appear to suffer. This was planning, not marking, but it is the kind of time that can go back into feedback.
What are the limits?
Accuracy is the weak point. In the Steiss study, accuracy was one of the qualities where humans did better. A confident but wrong comment can do more harm than none, especially in physics or maths where one sign error changes everything.
Most studies are on writing. Evidence for AI feedback on multi-step science and maths solutions is thinner. We cover how students can check AI comments on their working in getting useful AI feedback on step-by-step solutions.
Comparisons matter. "Better than nothing" and "as good as a teacher" are different claims. Read which one a study tested.
Feedback still has to be used. The research on formative assessment that Black and Wiliam made famous is about feedback students act on, not feedback they receive. Our explainer on Black and Wiliam's Inside the Black Box has the background.
A checklist for teachers trying AI feedback
Give it your rubric. Paste the marking criteria and one example of a strong answer. Being criteria-based was the one quality where AI did not trail humans in the Steiss study.
Ask for questions, not corrections. For example: "For each error, write one question that would help the student find it. Do not give the corrected answer." This mirrors the guiding questions that Tutor CoPilot encouraged.
Check every factual or numerical comment before it reaches a student. Read the AI's comment against the student's actual working, not just against the final answer.
Limit it to two or three points per piece of work, prioritised. A short list is easier for a student to act on.
Build in a response. Ask students to revise and mark what they changed, so you can see whether the feedback was used.
Protect student data. Remove names and identifying details before pasting work into any tool, and follow your school's policy.
Spot-check regularly. Every week, compare a few AI drafts with your own judgement and note where it goes wrong for your subject.
What this means for students and parents
If your school starts using AI feedback, ask the same question the studies ask: does a teacher check it, and do students revise after it? At home, students can practise the same habit with any AI tool by asking for hints rather than answers. Fermi's AI tutor notices where a student is stuck and decides whether to explain, ask or show.
