Guides · AI marking

Can you use ChatGPT to mark essays?

Plenty of teachers already do: paste in the mark scheme, paste in an essay, ask for a mark. It works well enough on one script to be convincing. The problems show up on the thirtieth, and in the data protection policy.

Updated 9 October 2026 · 8 min read

What a chatbot does well

Give a capable chatbot a clear mark scheme and a single essay and it will usually produce something reasonable: a level, a mark, a paragraph on strengths and a target. For a teacher looking at a pile of thirty, that is a compelling first experience, and it is why so many are trying it.

The trouble is that marking is not one essay. It is thirty essays marked to one standard, recorded, returned and defensible. That is where a general-purpose chat window starts to struggle.

Five ways it goes wrong across a class set

1. The standard drifts

Each new chat starts from nothing, and a long chat carries the influence of every essay before it. Paste the same script twice, in two conversations, and you may get two different marks. Over a class set, that means the mark depends partly on when a script was pasted, which is exactly the inconsistency marking is meant to avoid.

2. It marks too low

Asked to mark, language models often imagine an ideal answer and deduct for everything missing from it. That is the opposite of best-fit, levels-of-response marking, and it pulls marks down across the board, hardest on middle-band answers. It is the single most common failure in automated essay marking. See how GCSE essays are meant to be marked.

3. It quotes things the student never wrote

A chatbot asked for evidence will produce evidence. Sometimes that is a real phrase from the essay; sometimes it is a plausible paraphrase presented in quotation marks; occasionally it is invented outright. A student who is told their feedback rests on a sentence they did not write will, reasonably, stop trusting the feedback.

4. Score and level disagree

It is common to see a chatbot describe an answer as "Level 3" and award a mark that sits in Level 4's range, or add up criterion scores that do not match the total it reports. On one script you will spot it. On thirty, at ten at night, some will get through.

5. There is no workflow

Copying thirty essays in, copying thirty sets of feedback out, matching each to the right student and getting it back to them is a significant amount of clerical work. The marking may take seconds; everything around it does not.

The data protection question

Student essays are personal data. They carry a name, often more personal detail than you would expect, and they belong to a child. Before pasting them into any AI tool, the questions a school will ask are:

• Is this tool approved by the school? Many schools have a list. A personal account on a consumer chatbot is usually not on it.
• Is the work used to train models? Consumer chatbot accounts may use conversations to improve the service, depending on the provider and your settings.
• Where is it stored, and for how long?
• Is there a data processing agreement between the school and the provider?

Removing names before pasting reduces the risk but does not remove it, since the work itself can identify a student. Our page on whether AI marking is allowed covers what the Department for Education has said and what a head of department will want to know.

If you are going to do it anyway

If your school permits it, these steps make chatbot marking safer and more consistent:

What a dedicated marking tool adds

Most of the checks in that list are things a purpose-built tool can do automatically, every time, instead of relying on a tired teacher to do them by hand. That is the difference worth paying attention to when comparing tools, more than which AI model sits underneath.

In Paddle's case:

• One standard for the whole set. Every script is marked against the same scheme and instructions, not a conversation that drifts.
• Best fit is built into the marking rules, which forbid deducting from an ideal answer.
• Every quote is checked against the student's submitted text before you see it. If it is not there, the criterion is flagged.
• Every score is checked against its band, and any adjustment is shown to you, not made silently.
• No copying and pasting. Students submit through one link without an account; you review, edit and publish from one screen.
• Student work is stored in the EU and never used to train AI models, and schools can request a data processing agreement.

Common questions