What the research shows
The most useful study so far comes from researchers at UC Irvine and Arizona State, who had ChatGPT score about 1,800 history and English essays from middle and high school students, on a 1 to 6 scale, using the same rubric excerpt the human raters used. The paper, published in 2024 in Computers and Education: Artificial Intelligence, concluded that the scores were close enough to human scoring to be useful for low-stakes, formative purposes. The numbers behind that, as reported by the Hechinger Report, show where “close enough” ends.
Essays scored 1 to 6. Trained human raters gave exactly the same score as each other about half the time. Figures as reported while the paper was under review. Source: Hechinger Report, 2024, on the UC Irvine study.
Within one point sounds reassuring until you remember that one point on a six-point scale can be the difference between proficient and not. The study also found ChatGPT bunched its scores in the middle of the scale, between 2 and 5, so essays at the top and bottom tend to get pulled toward the middle, and those are often the ones where a grade matters most. The lead researcher’s advice was plain: not accurate enough for a high-stakes test or a grade that counts, fine for a preliminary score on a first draft.
That study used one model on one set of essays. Models have changed since, and your rubric and students are different, so treat it as a strong reason for caution rather than a measure of today’s tools.
Where AI helps, and where it should not
| Good uses | Keep for yourself |
|---|---|
| Drafting a rubric from your assignment and standards. | Any grade that goes in the gradebook. |
| Drafting criterion-by-criterion feedback on a draft. | Judgments about effort, honesty or AI use. |
| Turning your shorthand notes into clear comments. | Feedback on personal or sensitive writing. |
| Spotting common problems across a class set, without names. | Anything with a student’s name or record in it. |
1. Start from your rubric
AI feedback is only as good as the criteria it measures against. Without a rubric, a model falls back on what it thinks good writing is in general, which rarely matches what you taught. With your rubric, it can tie every comment to a criterion your students already know.
If you do not have one, AI is good at drafting it. Give it the assignment, the grade level and the standards you are assessing, and ask for four levels per criterion described in terms of what the student does, not adjectives like “excellent” or “weak”. The grading rubric template asks for exactly that, plus a checklist students can use on their own work before they hand it in. Then read it as a student would: if two levels sound the same, rewrite them before you use them.
2. Draft feedback, not grades
Asking a model to “grade this essay” invites the very use the research warns against, and the feedback that comes with it is usually general: “strong thesis, consider adding more evidence”. A prompt for feedback is different. It names the reader, a student who will revise, and asks for comments tied to quoted sentences, with no score.
Grade this essay.
{{essay}}I teach 9th-grade English. Below are the assignment, my rubric and one student's essay, with the name removed.
Assignment:
"""
{{assignment}}
"""
Rubric:
"""
{{rubric}}
"""
Essay:
"""
{{essay}}
"""
Write feedback the student will read before revising.
For each rubric criterion: one strength and one next step, each tied to a sentence quoted from the essay.
Then the two changes that would improve the essay most, in plain words a 14-year-old will follow.
Under 250 words, with a heading for each criterion.
Do not give a score or a grade; I will decide it. Do not rewrite the student's sentences for them.
Example of a next step: "Your claim in paragraph 2 needs a quote from the book to back it up."
A good result is feedback the student can act on without asking me what it means.
If the essay does not address a criterion at all, say so plainly instead of guessing.What changed. Quoted sentences make every comment checkable in seconds and show the student exactly where to look. “Do not rewrite the student’s sentences” keeps the thinking with the student, which is the point of revision. And leaving out the score avoids the problem of a number anchoring your own judgment.
Scores from the Prompt Grader’s free ten checks, run on October 28, 2026.
The feedback on student writing template is a fuller version: feedback for the student, plus a separate table of suggested rubric levels, with the evidence, for you to confirm or overrule. If you would rather decide before seeing any suggested level, delete that part. Save the one you settle on to My Prompts, so every essay in the set gets the same instructions; consistency across students matters more here than in almost any other use.
3. Review before students see it
Read each piece of feedback as the student will, and check three things. Every quoted sentence is really in the essay; models occasionally misquote, and a student who cannot find the sentence stops trusting the feedback. The next steps match what you taught; delete any advice that contradicts your instruction. And the tone fits the student; you know who needs encouragement and who needs a push, and the model does not.
Check consistency too. Before using a prompt on a whole class set, run it on two essays you have already marked, one strong and one weak, and compare its feedback with yours. If it praises the weak one or nitpicks the strong one, adjust the prompt before anyone sees it. The guide to testing prompts shows how to do this properly, and the Prompt Tester runs one prompt on several examples at once.
Feedback language can also carry bias in ways that are easy to miss, such as warmer words for some students and more about “effort” for others. The guide to fair performance reviews covers research on this in workplace writing, and the same check applies: every comment should point to something in the work.
4. Keep student data out
In the US, FERPA protects students’ education records, and many states and districts add their own rules about which apps may handle student work. The safe default is simple: use only AI tools your school or district has approved, and never paste a student’s name, ID, grades, accommodations or anything about their family or health. Remove the name from the essay before you paste it, and check the text itself for names of classmates or personal details. Safe Paste swaps names for placeholders in one step, and the anonymizing guide covers what else to look for. Personal writing, such as a reflective essay about family or health, is better kept out of AI tools entirely.
Telling students, and AI detectors
Be open about it. Students are more likely to take feedback seriously, and to be honest about their own AI use, when they know how yours works: “I use AI to help draft comments on your rubric, and I read and edit every one. Your grade is mine.” If your school has a policy on staff use of AI, follow it.
The reverse question, whether AI can tell you that a student used AI, has a clear answer: not reliably. OpenAI withdrew its own AI-writing classifier in 2023, citing its low rate of accuracy; at launch it caught 26% of AI-written text and wrongly flagged 9% of human writing. Treat any detector’s verdict as a reason for a conversation, never as evidence on its own.
For the rest of the job, the teachers collection has prompts for report card comments, replies to parents and quizzes from your own material, and the education use case covers planning and privacy across a whole school year.
A checklist to copy
Before you prompt [ ] An AI tool your school or district approves [ ] Student name and personal details removed from the essay [ ] Personal or sensitive writing kept out entirely [ ] Rubric ready, with levels that describe what students do In the prompt [ ] Grade level and the assignment [ ] Your rubric, pasted in full [ ] Feedback for the student, tied to quoted sentences [ ] "Do not give a score or a grade" [ ] "Do not rewrite the student's sentences" [ ] "Say so if a criterion is not addressed" Before students see it [ ] Every quote is really in the essay [ ] Advice matches what you taught [ ] Tone fits this student [ ] Prompt tested on two essays you already marked [ ] The grade is yours, decided from the work
Key takeaways
- In research, ChatGPT matched a trained grader’s exact essay score only about 40% of the time. Do not use it for grades that count.
- Use AI to draft rubric-based feedback on drafts, then review and edit it before students see it.
- Tie every comment to a quoted sentence and a rubric criterion, and leave the score out of the prompt.
- Use approved tools only, and remove names and personal details before you paste.
- AI detectors are not reliable evidence. Talk to the student instead.
Questions
Can ChatGPT grade essays accurately?
Not accurately enough for grades that count. In a study of about 1,800 secondary school essays, ChatGPT was within one point of trained human graders 76% to 89% of the time, depending on the batch, but gave exactly the same score only about 40% of the time, and it bunched scores in the middle of the scale. The researchers suggested it only for low-stakes uses, such as a preliminary score on a first draft.
Is it legal to put student work into ChatGPT?
It depends on your school’s rules, your state and the tool. FERPA protects student education records in the US, and many districts approve specific AI tools for staff. Use an approved tool, remove the student’s name and any personal details first, and keep personal writing out entirely. When in doubt, ask your school’s technology or privacy lead.
Can AI make a grading rubric?
Yes, and it is one of the most useful things it does for teachers. Give it the assignment, grade level and standards, and ask for criteria with three or four levels described by what the student does, not by adjectives. Then read every level as a student would and rewrite any two that sound the same. The grading rubric template is set up this way.
Should I tell students I use AI for feedback?
Yes. Say what the AI does, that you read and edit every comment, and that the grade is yours. It models the honest use you want from students, and it makes it easier to ask them to be open about their own use. Follow your school’s policy on staff use of AI if there is one.
Can AI detect if a student used AI?
Not reliably. OpenAI withdrew its own AI-writing classifier in 2023 because of its low accuracy, and other detectors also produce false positives, which can fall on students who did nothing wrong. Treat a detector’s result as a reason to talk to the student about their process, never as proof.
Prompts to use with this
- Give feedback on student writing against a rubricYou need specific, encouraging feedback tied to rubric criteria for a piece of student work.
- Build a grading rubric from an assignmentYou set an assignment or task and need a rubric that makes grading consistent and tells learners what good work looks like before they start.
- System prompt for an AI grader that scores answers against a rubricYou are evaluating an AI feature or a batch of written answers and need a grader prompt that scores consistently, shows its evidence and resists manipulation.
- Turn classroom observation notes into feedback for a teacherYou observed a lesson as a coach, mentor or school leader and need feedback that is specific, fair and focused on one next step.