LeasonAIGrading Writing With AI: The 4 Things Not to Do
Before you use AI to speed up writing feedback, read the four failure modes that get teachers into trouble — and the safer patterns that actually help students.
Grading Writing With AI: The 4 Things Not to Do
AI can genuinely accelerate writing feedback. It can also, if used carelessly, produce feedback that's worse than what students would get from a rushed teacher — because it's confidently wrong, culturally flat, or aimed at the wrong reader. This piece is the negative version of a workflow guide: four things not to do, and what to do instead.
1. Don't let AI assign the grade
The temptation is obvious: paste the essay, add the rubric, ask "what score does this get on each criterion?" The output looks plausible. It's also unreliable in ways that matter.
AI models score based on surface features they've seen correlated with quality in training data: length, vocabulary sophistication, sentence variety. They're measurably less accurate at scoring for the things rubrics actually target — argument structure, evidence quality, insight — and they're measurably biased against writing that deviates from standard academic English, which disproportionately affects multilingual students.
What to do instead: Use AI to describe the essay against the rubric ("this essay has three body paragraphs; the thesis is stated in paragraph one; the second paragraph's evidence is a quotation from the assigned text"), then you assign the score. Description is a much easier task and the AI is much better at it.
2. Don't send it back as-is
AI-generated feedback tends to be long, generic, and tone-deaf. A student who gets three paragraphs of "your thesis could be strengthened by considering more nuanced perspectives" learns nothing except that their teacher didn't read the essay.
What to do instead: Use the AI output as your first draft, then cut it hard. The rule I use: a feedback comment survives if it (a) names a specific line or paragraph, (b) says one concrete thing to do, and (c) sounds like something I would actually say. Everything else gets deleted. A three-paragraph AI response usually collapses to two sentences of comment. That's the goal.
3. Don't skip the read
The seductive workflow is "paste, generate, copy, paste to student." The problem is you never read the essay. When a student comes to office hours next week and asks about your feedback, you have nothing to say — you didn't read the piece, you read the AI's summary of the piece.
This erodes trust faster than any single bad grade. Students can tell when you haven't read their work.
What to do instead: Read the essay first. Then run the AI. Compare its noticing to yours. Keep what's useful, override what's wrong or missing. The AI is a second reader, not a replacement reader.
4. Don't use it on personal or high-stakes work
Some writing shouldn't go through a third party at all. This includes:
- Personal narratives and college essays.
- Reflections on sensitive topics (identity, family, trauma).
- IEP-related work where the student's specific challenges shouldn't be exposed to a general-purpose model.
- Final drafts of anything summative.
The reason isn't only privacy (though that matters). It's that AI feedback flattens voice. On a first argumentative essay, that's fine — voice isn't the point. On a college essay, voice is the point, and the AI will smooth it out.
What to do instead: Read this work yourself, unassisted. If you don't have time for that this week, extend the deadline or return it without written feedback and offer a 5-minute conference instead. A short honest conversation is better than a long AI-smoothed comment.
The pattern underneath the four rules
All four failure modes come from treating AI as a substitute for teacher judgment instead of a support for it. The safe framing is that AI does the mechanical parts — describing, listing, generating first drafts of comments — and you do the judgment parts — grading, prioritizing, deciding what's worth saying. When the tools are used that way, they're a real time-saver. When they're used the other way, they're a slow-motion trust problem.
A test I use: before I send feedback, can I defend every sentence in it as something I actually believe about this specific piece of writing? If yes, ship it. If no, cut whatever fails the test. That rule alone prevents most of the AI-feedback disasters.