Photograph a paper, drop it into ChatGPT and ask for a grade: the idea is tempting, and it sometimes works surprisingly well on a single paper. The trouble starts with a stack of thirty papers, a precise rubric to follow and papers that carry the names of students who are often minors. This guide describes a method that limits bad surprises, then the limits no prompt can get around.
What ChatGPT can do with a student paper
Recent chat assistants accept images: they can read a photographed handwritten paper, reproduce its content and compare it with what was expected. With legible handwriting and a well-posed question, the suggested grade is often plausible, and the comment that comes with it can save you writing time.
Keep in mind, though, that the tool was not designed for grading: it knows neither your rubric, nor your expectations, nor how you graded the previous papers, unless you give it all of that again every time. Everything that makes grading consistent therefore rests on you.
The method, step by step
1. Anonymize first
A student paper is personal data: first name, surname, class, sometimes even the handwriting can identify a student. Before taking the photo, cover the name (a sticky note, a strip of paper) or number the papers and keep the matching list on your side. It is the most often forgotten step, and the most important one.
2. Write the rubric down in black and white
The model will grade against what you give it, nothing more. A vague rubric (“5 points for question 2”) produces vague grades; a criterion-by-criterion rubric (“1 point for the method, 1 point for the calculation, 0.5 for the unit”) produces grades you can check. If you are not used to breaking things down, our guide on how to build a grading rubric offers a method and examples by subject.
3. Use a structured prompt
Attach the exam and, if you have one, the answer key: they serve as a reference for the expected results. Ask for a score per criterion rather than an overall grade, or you will not be able to tell where the gap with your own judgment comes from.
4. Check every suggestion
- Add up the total yourself: the stated sum does not always match the detailed points just above it.
- Reread the quoted passages: if the model misread a word or a number, the justification will show it.
- Compare with the first papers you graded yourself, to spot drift from one paper to the next.
- Keep the final say: the grade is yours, and you are the one accountable for it to the student and their family.
The limits on a whole batch
On one paper, the method above holds up. On thirty, several problems appear that have nothing to do with the quality of the prompt.
| Problem | What it means in practice |
|---|---|
| The rubric gets lost | In a long conversation, the instructions given at the start carry less and less weight; with one conversation per paper, you have to paste everything in again each time. |
| Consistency is not guaranteed | Two equivalent papers can get different grades depending on the order or the wording of the request. |
| Nothing is placed on the paper | The remarks stay in the conversation: you have to copy them by hand onto the paper the student gets back. |
| No grade sheet | Totals have to be entered one by one into your spreadsheet or gradebook. |
| Anonymization is manual | Each paper has to be masked before the photo, then matched back to its student. |
Taken together, this back and forth eats up a good part of the time the tool was supposed to save. That is what many teachers find after an enthusiastic first try on a single paper.
Student papers and data protection: what to check
Sending papers to an online service means sharing personal data about students, who are often minors. Before doing so, three questions deserve a clear answer.
- Are the papers used to train a model? On consumer assistants, it depends on the plan and the account settings: check the terms in force and turn off reuse if the option exists.
- Where is the data processed? Consumer services often process data outside the European Union. That is not forbidden in itself, but the transfer must be covered by proper safeguards (under the GDPR, for example).
- Who is responsible? Your school or district usually has a data protection officer: that is the person to consult before adopting a tool for your classes.
In every case, a paper anonymized before it is sent greatly reduces the risk: a text with no name, class or school says much less about a student.
The alternative: a tool built for grading papers
Correxio takes the same idea, an AI that reads the handwritten paper and suggests a grade, but organizes it around grading a whole batch:
- the rubric is entered once for the whole batch, criterion by criterion, and the AI can draft a first version from the exam;
- the student's name is hidden automatically before the AI reads the paper, and the data is hosted in France (security and GDPR);
- each paper gets a score per criterion with its justification and annotations placed on the scan, right where they apply;
- you review and adjust every grade before exporting the grades as CSV and the annotated papers as PDF, ready to hand back.
In short
ChatGPT can help grade a single paper, provided you anonymize it, supply a detailed rubric and check everything. For a whole batch, consistency, annotations and export call for a tool designed for grading. Either way, the final grade remains the teacher's.