AI marking: the complete guide for teachers

What is AI marking?

AI marking is using artificial intelligence — in 2026, almost always a large language model — to assess student work against criteria and produce marks, comments, or feedback. The term covers two jobs that look similar and behave very differently.

The first is marking as a verdict: a mark or a grade against a mark scheme, the way an exam script is marked. The second is marking as feedback: strengths and next steps against curriculum criteria, drafted for a teacher to review, edit, and deliver. A wrong verdict gets recorded; a wrong draft gets caught by the teacher reading it. The evidence, the rules, and the risk all divide along that line, which is why it is worth keeping the two apart from the start — the full distinction is on AI writing feedback vs AI marking.

This guide covers both, because teachers searching for “AI marking” usually mean the whole question: can a machine read my class's books, say something useful about each piece, and give me my Sunday back — without getting me in trouble with the data protection officer or the moderator?

Does AI marking actually work?

The published evidence is better than sceptics expect and narrower than the marketing suggests. In the largest UK trial — No More Marking's 2025 comparative judgement assessment of around 70,000 pieces of Year 7–9 writing across 177 schools — AI agreed with teachers' judgements 83% of the time, which matched the level at which human judges typically agree with each other. Vendor studies on GCSE mark schemes report similar or higher agreement.

The honest caveats: almost all of that evidence is secondary-level; agreement on a ranking is not the same as writing the right next step for a nine-year-old; and AI remains weakest on nuance — deliberate rule-breaking, humour, a voice that doesn't fit the rubric. The full evidence, including what none of the studies measure, is on is AI marking accurate?

Is AI marking allowed?

In England, yes — with two rules that matter. The DfE's policy paper on generative AI in education lists feedback among the uses it sees benefit in, and its training materials name “marking work, giving feedback” as a legitimate use. The first rule is oversight: any AI-produced content “requires critical judgement to check for appropriateness and accuracy”, and final responsibility stays with the teacher and the school. The second is a bright line: AI shouldn't be used to make summative judgements or predictions about student attainment. AI can inform a teacher's judgement; it cannot replace it. What the DfE documents actually say, quoted and linked, is on the DfE guidance explainer.

At Year 6, the STA's KS2 teacher assessment guidance adds the independence rules — and they cut both ways. A teacher using AI to inform their own assessment of a completed piece does not change how that piece was produced: if the pupil wrote it independently, it remains independent evidence. But writing a pupil has edited or rewritten after seeing AI feedback is not independent, and moderators must be told where AI feedback was used so that work can be excluded. Keep originals in the moderation collection; treat post-feedback redrafts as taught work. The exact STA wording and a practical dos-and-don'ts list are on AI marking and independent writing.

If your school hasn't written any of this down yet, there's a free AI policy template for primary schools with a statutory writing-assessment annex built on exactly these rules.

Is AI marking safe for pupil data?

This is where general chatbots and purpose-built tools part company. The DfE recommends that personal data is not put into generative AI tools, and where it is strictly necessary, the school must ensure the product complies with data protection legislation and its own privacy policies. Its training materials add two points teachers sometimes miss: never input details that could identify individual students unless advised it is safe, and students own the copyright to their own work.

The practical checklist — what to ask any vendor, what UK GDPR requires, and how Howay handles children's writing — is on is AI marking safe?

Can AI mark handwritten work?

Yes — and at primary this is the whole game, because the writing that needs marking lives in exercise books, not documents. Modern vision models can read a child's handwriting from a photograph well enough to assess the writing itself: no typing, no scanning hardware, no asking Year 3 to transcribe their own story. The teacher photographs the completed page and reviews what comes back — including whether the handwriting was read correctly. How that works, and where it still struggles, is on AI marking for handwritten primary writing.

How do teachers use AI for marking in practice?

The workflow that fits the rules above has four steps. The class writes as normal, on paper. The teacher photographs the completed work. The AI drafts feedback against the curriculum criteria the teacher chose. The teacher reviews and edits every word before anything reaches a child, a parent, or a book. The piece of writing itself is untouched — which is what keeps the evidence trail clean for moderation.

The main fork in the road is the tool. General chatbots are free and flexible, but the DfE's own guidance flags the catch: LLMs are designed to vary their answers, so the same piece of work can get different feedback on different days, and you carry the data protection question yourself every time you paste. Purpose-built marking tools exist to fix consistency, hold the curriculum criteria for you, and put the data handling on a compliant footing. The honest comparison is on Howay vs ChatGPT for marking, and the wider market on the best AI writing feedback tools for UK primary.

Where Howay fits

Sources

1. DfE policy paper, “Generative artificial intelligence (AI) in education” (updated August 2025) — the DfE's formal position on AI for feedback, teacher oversight, and pupil personal data.
2. DfE support materials, “Use cases of generative AI in education” Module 4 — marking and feedback use cases, the summative-judgement rule, and the personal-data and copyright warnings.
3. STA, “Key stage 2 teacher assessment guidance” (updated March 2026) — the independence rules for writing evidence, including writing edited or rewritten following AI feedback.
4. Christodoulou, D. (2025), “AI is uncannily good at judging writing”, No More Marking — c.70,000 pieces of Year 7–9 writing, 177 UK secondary schools; AI agreed with 83% of 133,983 human comparative judgements.

Common questions

What is AI marking?

AI marking is using artificial intelligence — usually a large language model — to assess a piece of student work against criteria and produce marks, comments, or feedback. In practice it covers two quite different jobs: producing a verdict (a mark or grade against a mark scheme) and drafting formative feedback (strengths and next steps against curriculum criteria) for a teacher to review. The evidence, the rules, and the risks are different for each, which is why the distinction matters more than the label.

Is AI marking allowed in UK schools?

Yes, with conditions. The DfE's generative AI policy paper lists feedback among the uses it sees benefits in, and its training materials name 'marking work, giving feedback' as a legitimate use — provided the teacher checks outputs for accuracy and appropriateness and final responsibility stays with the professional. The same materials are clear that AI should not be used to make summative judgements about student attainment, and the STA's KS2 guidance sets specific rules about AI feedback and independent writing for moderation.

Is AI marking accurate?

On the best-evidenced task — comparative judgement of writing — AI now agrees with teachers about as often as teachers agree with each other: 83% agreement across roughly 70,000 scripts in No More Marking's 2025 assessment. But almost all published evidence is secondary-level, and agreement on a ranking is not the same as writing the right next step for a nine-year-old. The practical question is who catches the errors, which is why teacher review before anything reaches a child matters more than any headline percentage.

Can AI mark handwritten primary writing?

Yes — modern vision models can read children's handwriting well enough to assess the writing on the page, without typing or transcription. This matters at primary because almost all everyday writing is handwritten in books. A teacher photographs the completed page, the AI extracts the writing and assesses it against the criteria the teacher chose, and the teacher reviews the result. Accuracy of the handwriting reading should be checked the same way as accuracy of the feedback: the teacher sees both before a child does.

Does AI marking count against independent writing at KS2?

The question splits in two. A teacher using AI to inform their own assessment of a completed piece does not change how that piece was produced — if the pupil wrote it independently, it remains independent evidence. But the STA guidance is explicit that writing a pupil has edited or rewritten following feedback from AI is not independent, and writing produced in full or in part by AI must never be used for teacher assessment judgements. So keep original versions in the moderation collection, and make clear to moderators where AI feedback was used.

Can I just use ChatGPT for marking?

You can, and many teachers have tried it. The trade-offs are real: the DfE's guidance notes that general-purpose LLMs are designed to vary their answers, so the same work can get different feedback on different days; you have to build and re-supply your own criteria each time; and pasting pupil work into a general chatbot raises data protection and copyright questions the DfE materials specifically warn about. Purpose-built tools exist to fix exactly those three problems — consistency, curriculum criteria, and a compliant pipeline.