How to Mark Student Work Using ChatGPT

Most teachers like reading their students' thinking. What wears people down is the volume of it - the stack of scripts on a Sunday, the same three comments written thirty times, the rubric held in your head while your attention drifts on script twenty-two. A general-purpose chatbot can take a real bite out of that, if you set it up properly.

This guide covers how to use ChatGPT for marking: getting started, the prompts that actually work, what to check before feedback reaches a student, and - honestly - where a chatbot stops being the right tool.

What ChatGPT can and cannot do for marking

ChatGPT is a general-purpose language model. Given a question, a mark scheme and a student's answer, it will produce a mark and a written rationale, and it is genuinely good at the repetitive part: spotting whether a required point is present, drafting a comment, keeping a consistent tone across a class set.

What it does not do is hold anything for you. It has no memory of your mark scheme between sessions, no notion of your department's standard, and no way to tell you which criterion earned which mark unless you engineer that into the prompt yourself. It is an assistant you brief from scratch every time - which is fine, as long as you know that is the deal.

Getting started

There is no trick to it. Open ChatGPT, spend twenty minutes experimenting on work you have already marked, and compare what it gives you against what you gave. That comparison is the whole calibration exercise - it tells you where the model is generous, where it is harsh, and what your prompt is failing to say.

Have three things ready before you start:

• The question, exactly as the students saw it.
• The mark scheme, including where each mark is earned.
• A student answer, typed or pasted verbatim.

A minimal prompt that works:

You are an experienced GCSE Science teacher. Below is a question, a marking guide and a student answer. Give a mark, a breakdown of how you reached it, and feedback for the student.

Question: Explain how animal cells and plant cells differ. (4 marks)

Marking guide: Plant cells are rectangular, have a cell wall, chloroplasts and one large vacuole. Animal cells are circular (1 mark), have no cell wall (1 mark), no chloroplasts (1 mark), and many small vacuoles (1 mark).

Student answer: [paste verbatim]

Writing prompts that produce useful feedback

The difference between feedback that helps and feedback that reads like a horoscope is almost entirely in the prompt. Four things make the difference:

• Be specific about where marks live. Paste the mark scheme in full. A model given "mark this out of 4" will invent its own criteria, and they will not be yours.
• Show it a good and a bad example. One of each is enough to establish what you mean by useful.
• Constrain tone and length. "80–120 words, encouraging, plain English, no more than two next steps" produces a comment a fourteen-year-old will read.
• Require evidence. Ask it to quote the student's own wording for every point it makes. This is the single most effective instruction - it forces the model to stay anchored to what is actually on the page.

A fuller version, for an essay:

Role: You are an experienced GCSE English teacher marking a student essay.

Apply the rubric below. Give a total mark, a breakdown across the criteria, and feedback for the student that is specific (names what they did, with a quotation), actionable (says how to improve next time) and encouraging.

Rubric (20 marks): Structure (5) - clear introduction, organised paragraphs, coherent conclusion. Content and relevance (10) - accurate, relevant examples, answers the question. Grammar and mechanics (5) - spelling, punctuation, sentence structure.

Keep feedback to 120 words and end with two next steps.

Student essay: [paste verbatim]

Always check the output

Treat every mark as a draft. That is not a disclaimer - it is how the tool should be used.

• Read the rationale, not just the number. A right mark reached by wrong reasoning will mislead the student and will not survive a parent asking why.
• Watch the extremes. Models tend to bunch marks in the middle; the strongest and weakest scripts in a class set are where you will most often disagree.
• Add what only you know. That this student has been working on paragraphing for a month is not in the answer, and no model can infer it.

Student data, GDPR and your school's policy

This is the part to get right before you paste anything. The guidance is straightforward:

• Never put identifying details into a general-purpose chatbot - no names, no candidate numbers, no dates of birth, nothing in the answer itself that identifies the student.
• Pseudonymise. "Student A" marks exactly as well as a real name.
• Check your school's AI policy first, and your data protection lead's position on it. Policies vary widely and the one that matters is yours.

The Department for Education published guidance on generative AI in education covering exactly this. The short version: AI can help with planning, marking and admin, but it must not make unsupervised decisions about learners, and personal data needs a lawful and secure route.

Where a general-purpose chatbot runs out of road

For a handful of scripts, ChatGPT is fine. The friction shows up at scale, and it is structural rather than a matter of quality:

• You re-supply everything every time. The mark scheme, the tone, the examples - none of it persists, so the setup cost repeats on every batch.
• Nothing links a mark to a criterion. You can ask for a breakdown, but the model is composing one, not reporting one, and there is no audit trail to hand a moderator.
• Multi-question papers get unwieldy. A full exam with a long mark scheme and thirty scripts is a lot of pasting, and consistency across the batch is on you to check.

That is the gap purpose-built tools exist to close. Marking.ai marks to your own marking guidelines and mark scheme criterion by criterion, and shows the evidence behind every mark - and a teacher reviews and approves every mark before it reaches a student. A faster first pass, not an unchecked final mark. It is free to try: 20 submissions, one-time, no card required to start.

The point of any of this

You did not go into teaching to write "good use of evidence - develop your analysis" for the twenty-ninth time. Whether you use a chatbot with a well-built prompt or a tool made for the job, the aim is the same: get the repetitive first pass off your evening, and keep the judgement that actually needs you.