Feedback students actually use

Why the feedback is so nice

Ask a model to comment on a student’s draft and you get warmth. “What a thoughtful piece.” “You’ve made a great start here.” “Keep going.” Then, some way down, a suggestion. Then more warmth.

The mechanism is worth knowing. Models are refined by having people rate their responses, and people rate encouraging, agreeable responses higher than blunt ones. The model learned that praise scores well. This is the same tendency that makes it agree with you when you push back on a correct answer, and it has a name, sycophancy, and a consequence for feedback: the useful part is buried in the part the student skips.

What the research on feedback says

Three findings hold up across many studies. Feedback that names a specific next action is used; feedback that describes quality in general terms is not. Feedback that arrives while the student can still act on it does far more than the same feedback after the grade. And a grade next to a comment makes the comment invisible: students read the number and stop.

So the target is: specific, actionable, timely, and separate from any mark. The model can produce the first two at scale if told to, and it is the only thing that makes the third possible for sixty students.

The prompt

Here is a Class 9 student's draft paragraph and the rubric.
Give exactly two pieces of feedback. Each must name a specific place in the draft (quote three or four words from it), say what to change, and say why it will improve the paragraph against the rubric.
No praise. No summary of what the paragraph does well. No closing remark.
Write it to the student, in plain sentences, under 80 words in total.

“No praise” is the instruction that fights the training. It works for the first several drafts and drifts, like every other constraint; restate it when the compliments creep back.

“Exactly two” is the other important number. Six pieces of feedback are a list the student does not act on. Two is a task. The best feedback teachers give is often one sentence, and the model can be held to that.

Feed-forward, not feedback

The most useful form points at the next piece of work rather than the last: “In your next paragraph, put the quotation before the claim and see whether it reads more convincingly.” The draft is over; the next draft is where the change happens. Ask for it explicitly: “Phrase each point as something to do in the next version, not as a description of what is wrong with this one.”

Comment banks, and choosing from them

For a class set, generate the feedback in bulk and then choose. Paste ten drafts with a code instead of names; ask for two points each. Read all twenty. You will find that around half are right, a quarter are right but not the most important thing, and a quarter miss the point. Keep the half, rewrite the quarter, replace the rest.

That reading takes fifteen minutes for a class of thirty and it is not optional. It is also where you notice the pattern (eleven students have the same problem with topic sentences) which is next week’s lesson and something no individual comment tells you.

The student who does not read comments

Written feedback is only useful if it is read, and many students do not read it. Two things that help: return the feedback without a mark, so the comment is the only thing on the page; and require a response (“Write one sentence saying what you will do differently next time”) before the next task begins. Both cost nothing and both raise the proportion of feedback that becomes action.

Where AI feedback should stop

The model can point at the draft. It cannot know that this student was told about topic sentences three times already and needs a different approach, or that this one has just started writing in paragraphs at all and the right feedback is “you did it”. Those judgements are about the person and they are yours. Use the model’s two points as a draft of your comment, and change them where you know something it does not.

The honest limit

Fast, specific feedback in volume changes what students expect. A class used to two precise points on every draft within a day will, reasonably, want the same on the pieces you mark by hand, and you may not be able to deliver it. Decide which pieces get model-drafted feedback and which get yours alone, say so, and make sure the ones that get yours are the ones where the relationship matters. Feedback is partly information and partly attention, and only one of those can be generated.