Feedback Without Rewriting: The Hardest Move in AI Tutoring
An AI that rewrites the student's essay teaches nothing. The skilled move is rubric-based feedback that diagnoses without doing the work — preserving the struggle where the learning actually lives.
Paste a rough paragraph into any general assistant and watch what happens. It hands back a cleaner paragraph. Tighter, better-ordered, the comma splices gone, a limp verb swapped for a sharp one. It looks like help. It is the single most destructive thing an AI can do to someone who is trying to learn to write.
Because the student didn't learn to write. The model did the writing. The gap between the draft they produced and the draft they got back is exactly the learning that never happened — and worse, it now wears the student's name. This is the writing-shaped version of the argument I made in The AI Tutor That Refuses the Answer: the moment the assistant supplies the finished thing, it suppresses the struggle where the skill is actually built. In math the finished thing is the answer. In writing it's the rewritten sentence. Same trap, different clothes.
The rewrite trap is a helpfulness instinct, not a bug
Nobody trained the model to be lazy. The opposite — it rewrites because it's trying to be maximally useful, and a better paragraph is the most legible proof of usefulness it can offer. You asked it to help with your essay; it improved your essay. Request satisfied.
That instinct is correct for a colleague and catastrophic for a learner. When I hand a draft to an editor, I want the fix — I already know how to write, and the edit ships. When a student hands a draft to a tutor, the fix is the one thing that must not happen, because the point was never the paragraph. The point was the student's growing ability to see what's wrong with their own paragraphs. Rewrite it and you've optimized the artifact while starving the writer.
There's a quieter theft underneath the pedagogical one. A rewritten essay is no longer the student's essay. Their voice — the slightly odd phrasings, the rhythm that's theirs — gets sanded into the model's default register: fluent, competent, and completely anonymous. You can feel it when you read a class set of AI-"helped" essays. They converge. Authorship is a thing you can quietly take from someone by being helpful at the wrong altitude.
What non-rewriting feedback actually looks like
The alternative isn't to be stingy or vague. "Make this stronger" is useless. "Nice work!" is worse. Good feedback is more specific than a rewrite, not less — it just stops one step short of doing the work. It names the weakness, explains why it's a weakness, locates it precisely, and points at the move that fixes it. Then it hands the pen back.
Take one flabby sentence: "There are many reasons why the policy was bad and it affected a lot of people in negative ways." Here's the same problem handled three ways:
| Response | What it does | What the student learns |
|---|---|---|
| Rewrite: "The 1834 Poor Law immiserated the rural laborers it claimed to protect." | Produces a great sentence | Nothing — and the sentence isn't theirs |
| Vague: "Try to be more specific and concise here." | Names a direction | Little — "specific" is abstract until you can see it |
| Diagnose + point: "This sentence makes two claims — the policy was bad, and it hurt people — but names no policy, no people, and no harm. Which policy? Who specifically? What did 'negative ways' actually mean for them? Rewrite it naming all three." | Diagnoses, locates, prescribes the move | How to spot and repair empty abstraction — a skill they keep |
The third response is longer and harder to generate than the rewrite. That's the tell. Rewriting is the low-effort move dressed up as generosity; diagnosis is the high-effort move that looks like it's doing less. A tool built for teaching, like an essay-feedback skill, is defined by its willingness to stay in that third column even when the student is begging for the first.
Rubrics turn taste into something a model can defend
"That's diagnosis by vibes," you might object — and you'd be right to worry. Ungrounded feedback drifts. The same essay gets praised on Monday and shredded on Tuesday because the model is reacting to surface fluency instead of measuring against anything. The fix is a rubric: an explicit set of criteria the feedback is anchored to.
A rubric does three things at once. It makes feedback consistent, because every draft is judged against the same dimensions. It makes feedback teachable, because the student learns the criteria and can eventually self-assess against them. And it makes feedback honest, because the model has to point at the criterion it's invoking — "your thesis is contestable but your body paragraphs never actually contest it" beats "the argument feels weak."
Feedback without a rubric is an opinion. Feedback against a rubric is a diagnosis — and a diagnosis is something a learner can act on, argue with, and eventually internalize until they no longer need you.
This is where the design work lives. A good rubric-driven skill separates the layers that a general assistant smears together. Correctness and mechanics — comma splices, agreement, a misused semicolon — are genuinely fine to fix outright, which is what a focused proofreading pass is for; nobody's higher-order writing ability is damaged by learning that its/it's rule once. But argument, structure, evidence, and voice are exactly the dimensions where the rewrite must never happen. The skill's whole intelligence is in knowing which layer it's touching and refusing to "help" on the wrong one. Mechanics: fix. Meaning: point.
Feedback tells them where they were; feedforward tells them where to go
There's a distinction worth stealing from writing pedagogy: feedback versus feedforward. Feedback is retrospective — here's what this draft did. Feedforward is prospective — here's the move to carry into the next paragraph, the next essay, the next assignment. The rewrite trap is pure feedback with the learning removed; it evaluates the artifact and stops.
The most valuable thing an AI writing tutor can do is convert every local note into a portable rule:
- Name the pattern, not just the instance. Not "this transition is abrupt" but "you tend to jump between ideas without a bridge — watch for it whenever a paragraph opens with a new subject." One instance is a correction; a named pattern is a skill.
- Prescribe the move, withhold the execution. "Open this paragraph with the claim, then the evidence" tells the student what to do without doing it. They still have to write the sentence — and writing it is the rep.
- Tie the note to the criterion. "This weakens your analysis score because you're describing the source, not interrogating it" teaches the rubric while fixing the draft, so the student gets better at seeing, not just at this essay.
- Escalate ownership over time. Early on, point precisely. Later, ask "which of your body paragraphs is weakest, and why?" A tool like an academic-writing-assistant earns its keep by handing the diagnostic job back to the student as they grow into it — the tutor's job is to make itself progressively unnecessary.
Do this and the student stops needing you for the thing you were pointing at, which is the only definition of teaching that means anything.
The discipline is in what you don't write
None of this is hard because the model can't do it. It's hard because the restrained response is more work to produce and less impressive to receive. A rewrite dazzles instantly; a good diagnosis asks the student to go do something. Every incentive — the user's "just fix it," the model's helpfulness training, the visible polish of a clean paragraph — pulls toward picking up the pen. The skill is in setting it down.
So the bar for an AI writing tutor is almost perverse: judge it by the sentences it refuses to write for you. An essay full of the model's prose and the student's name on top is a failure that looks like success. A messier draft the student fixed themselves, guided by feedback specific enough to act on, is the real thing — their argument, their voice, their growing eye. The whole craft of these tools, browsable among the AI tutoring skills, reduces to one stubborn move: diagnose relentlessly, point precisely, and never, ever pick up the pen.
Part 4 of the AI Tutoring series. Previously: Spaced Repetition Is the One Ed-Tech Idea That Survived. Next: The Assessment Loop: Use AI to Write the Test, Not Take It. Browse the AI tutoring skills or more builder insights.