Can ChatGPT grade my essay with a rubric?
9 min read · updated
ChatGPT can read your essay against a rubric you paste in and estimate a level for each criterion, but its scores are a rough second opinion and not a grade. In a 2025 preprint on 67 Italian university essays, agreement with human raters was low and repeated runs agreed only weakly with each other. Ask for a quoted sentence behind every score, check each quote and score yourself, and confirm that your course allows AI tools. Your instructor decides the grade.
What does it mean to ask ChatGPT to grade an essay with a rubric?
Asking ChatGPT to grade an essay with a rubric means giving a general-purpose chatbot your draft and the assignment’s rubric and asking it to pick a level or score for each criterion. What comes back is a prediction from a language model. It is not how your instructor will read the paper, and it is not a grade.
A rubric is a written standard, and a chatbot can read one. Reading is not judging, and this guide is about the gap. You get a prompt to try, a table for checking whatever the chatbot says, and an invented answer that looks fine until you check it.
If you want the manual version first, start with how to grade your own essay with a rubric. For how ChatGPT compares with other kinds of graders, see what AI essay graders can and cannot check.
Is ChatGPT’s grade accurate?
The research is mixed, and the study closest to a student’s situation is not encouraging.
Gaggioli and colleagues posted a preprint on arXiv in August 2025 that tested five language models, GPT-4 among them, on 67 Italian-language essays from a university psychology course. The essays were scored on a four-criterion rubric by human raters and by each model, three times per model. The authors report that agreement with the human raters was “consistently low and non-significant”, and that scores from repeated runs of the same model agreed only weakly (median Kendall’s W below .30). The models also tended to give higher Coherence scores than the humans did, and they were least stable and least consistent with each other on the context-dependent criteria, Pertinence and Feasibility.
The authors name the limits themselves: one discipline, one language, two human raters whose individual scores could not be separated, and a rubric built around interpretive criteria. Different results “may emerge with more objective, surface-level dimensions.” The models may also be older than whatever ChatGPT runs today, so do not assume the numbers carry over.
One other study is more favorable. A 2025 study of AP Chinese exam writing by Song, Lee and Jiao, with seven AI raters, found reasonable consistency under certain conditions, especially on story narration, though human raters were more reliable overall. Composite scores that combined human and AI raters were more reliable, which the authors say suggests hybrid scoring may help.
So how well ChatGPT scores depends on the task, the criterion and the run, and the human raters remain the reference point. For a student that points to a narrow use: let ChatGPT show you where to look, and make the judgment yourself.
One more caution comes from a 2025 paper on why language models hallucinate by Kalai, Nachum, Vempala and Zhang. The authors argue that models are trained and evaluated in ways that reward guessing over acknowledging uncertainty. For grading, that means a confident level with a fluent reason does not show the model found support in your text. Hence the quoted sentence in the prompt below.
Should you use ChatGPT on your essay at all?
First check your course. The UNC Writing Center’s handout on generative AI tells students to consult each syllabus for guidance or requirements on AI tools, and says you are always fully responsible for whatever writing you turn in. It also warns that these tools are not authoritative sources and can fabricate sources or credit words to the wrong author. Metro State’s guide to generative AI and writing support says to check the syllabus or ask your instructor, and to watch for sources that do not exist.
So before you paste anything:
- Read the syllabus section on AI tools. If it is silent or unclear, ask your instructor.
- Keep out any text you do not have the right to share, such as classmates’ drafts, restricted course documents and patient or client data.
- Check what the tool does with pasted text. Settings and policies differ by product and by account type, and they change. Metro State also advises caution with third-party tools because of privacy, and points to the tools your institution has approved.
Asking ChatGPT only to estimate and explain is a narrower use than letting it draft your sentences, but your course may restrict both.
How do you ask ChatGPT to grade against a rubric?
Paste the full rubric with every level, paste the essay, and ask for a level per criterion with a quoted sentence behind it. This template is one we wrote. We have not tested whether it produces reliable scores, and it does not change any of the limits above.
Below are an assignment rubric and my essay draft.
For each criterion in the rubric, give me:
1. The performance level that best matches my draft.
2. One sentence quoted exactly from my essay as evidence. If no
sentence fits, write "not enough evidence" and do not guess.
3. The words in the level descriptor that this sentence meets or misses.
4. What a reader would still need to see to place the draft one
level higher.
Do not give an overall grade. Do not rewrite my essay or suggest
replacement sentences. Do not comment on anything the rubric does
not mention.
RUBRIC:
[paste every criterion with all of its levels]
ESSAY:
[paste your draft]
Pasting every level matters because the bare name “Analysis” leaves the model to supply its own idea of analysis, while the descriptors are the standard your instructor set. An exact quote is something you can search for in your document; a summary of what the essay “seems to” do is not. “Not enough evidence” gives the model a way to decline, so it is less pushed toward a confident guess. An overall grade would be one number hiding which criteria it rests on. And the ban on rewriting keeps the writing yours, and keeps you within the rules of courses that limit AI-written text.
If the rubric is vague on a point, do not let the model settle it. Check the assignment sheet or ask the instructor. The UNC Writing Center’s Understanding assignments handout suggests exactly that for anything you do not understand.
How do you check what ChatGPT tells you?
Treat each criterion in the answer as a claim to verify. Run these checks in order; a criterion that fails any of them is not usable yet.
| Check | What to do | A warning sign |
|---|---|---|
| Quote exists | Search your document for the quoted sentence, word for word | The sentence is not in your essay, or it is a reworded version of one |
| Quote fits | Read the quote next to the criterion | It is about the topic but not about what the criterion measures |
| Descriptor match | Find the descriptor words the answer cites and read the neighboring levels | The reason uses general praise (“clear and engaging”) with no descriptor wording |
| Repeat run | Open a new chat, paste the same prompt and compare the levels | A criterion moves by a level or more between runs |
| Your own score | Grade the criterion yourself first, using the self-grading steps, then compare | You disagree, and neither of you can point to the sentence that settles it |
| Facts and sources | Verify claims and citations in your draft from the sources themselves | The chatbot says a source supports your claim, and you have not opened the source |
When ChatGPT and you disagree on a level, you have found a paragraph worth rereading. Go back to the descriptor and decide from the page.
What does a checked answer look like?
The material below is invented. The rubric wording, the essay lines and the chatbot answer are all made up for illustration, and the answer is not real ChatGPT output.
Illustrative criterion: “Uses relevant evidence to support claims.” Levels: Strong (the claim is supported by specific, relevant evidence that the writer explains); Adequate (relevant evidence, with little explanation); Weak (evidence is general or unrelated).
Illustrative essay lines: “Students need quiet places to study at night. A campus survey found that 62% of respondents had trouble finding a seat after 9 p.m. The library should extend its hours.”
Illustrative chatbot answer: “Strong. Evidence: ‘A campus survey found that most students struggle to find seats at night.’ The writer uses a specific survey and explains its relevance.”
Now run the table:
| Check | Result |
|---|---|
| Quote exists | Fails. The essay says “62% of respondents had trouble finding a seat after 9 p.m.”, not “most students struggle”. The model paraphrased and presented it as a quote. |
| Descriptor match | Fails. The “Strong” level requires that the writer explains the evidence, and the answer claims an explanation that is not in the three lines. |
| Repeat run | A second chat returns “Adequate”. The levels differ, so the criterion is not settled. |
| Your own score | After checking, you decide the evidence is specific, but nothing links it to the claim about hours. You place it at Adequate and write the missing explanation sentence as the next revision. |
You only see the missing explanation once you set the quote beside the page. To test whether the evidence fits the claim, check your evidence against the rubric one claim at a time.
What mistakes make a ChatGPT grade misleading?
- Taking the first answer. One run is one sample, so compare two before you rely on a level.
- Re-asking until the score goes up. A higher number on the fifth try measures your persistence. Settle disagreements with the descriptor.
- Pasting only part of the rubric. Without the lower levels you cannot see what separates adjacent scores.
- Ignoring weights. A rubric may count some criteria more than others, and a chatbot’s overall impression ignores that unless you apply the weights yourself. A revision priority table helps you decide what to fix first.
- Pasting its rewrite into your paper. That can break your course’s rules, and it can carry the model’s errors into your work.
- Treating fluency as evidence. A smooth explanation can describe a paragraph that is not in your essay.
Checklist before you rely on a ChatGPT answer
Where does a purpose-built checker differ?
A chat window gives you whatever format the conversation produces. A tool built for one job can fix the format. OnRubric’s rubric checker takes a rubric and a draft and returns an estimated score range for each criterion, along with sentences quoted from your draft and a list of what is missing. It checks each quote against your text in code, and an unverified quote is flagged with a question mark. That addresses one failure in the table above, the invented quotation, but not the larger limit. The result is still an estimate: it does not know your instructor’s standard, you decide what to revise, and your instructor decides the grade.
You can check a paper to see the format. The methodology page shows how far its estimates have missed human scores on test essays, and the privacy page explains what happens to the text you upload. As with ChatGPT, check your course’s rules before you upload a paper or a rubric.
Before you submit
Take the criterion where ChatGPT and your own reading disagreed most. Open the paragraph, read the descriptor and decide from the page. Make one revision there, then reread the paragraph against the rubric. Do not accept or reject the model’s answer on the strength of how sure it sounded.