AI can draft the marking. It cannot own the judgement.

Yes, a teacher in England may use AI to help mark work. Department for Education guidance is clear that generative AI can support low-stakes marking, such as homework and quizzes, and equally clear that it must not make summative judgements about a pupil's attainment. The teacher remains accountable for the professional judgement, the safeguarding, the data protection and the curriculum quality. Everything worth arguing about sits in the gap between those two sentences.

Is a teacher allowed to use AI for marking?

For low-stakes work, yes, with the school checking the output. For anything that contributes to a formal judgement about how well a child is doing, no. The distinction is not about how good the model is. It is about who is accountable. A grade is a statement a school makes about a child, and a school cannot delegate a statement it has to stand behind to a system that cannot be held to account for it.

That is a useful test to carry into any product demonstration. Ask where the tool sits on that line, and ask what it does to stop a teacher drifting across it on a tired Sunday evening. A tool that quietly produces something that looks like a grade has put the teacher on the wrong side of the guidance without ever telling them.

What is the difference between drafting feedback and judging a pupil?

Drafting is producing the words. Judging is deciding what is true about a child's learning. The first is a writing task and the second is a professional one, and they have been tangled together for so long that separating them feels strange.

In practice the separation is ordinary. A teacher who has read a set of books already knows that eight pupils have not grasped subordinate clauses. What takes the evening is writing that fact down twenty-eight times in a way each child can act on. A tool that drafts those twenty-eight comments from the teacher's own marking, in the teacher's own phrasing, for the teacher to edit and approve, has taken the writing and left the judgement exactly where it belongs.

Where does the pupil's work go when AI marks it?

This is the question most marking tools answer badly, usually by not answering it. Most AI works by sending the text somewhere else to be processed. For marking, the text is a named child's work, which is personal data about a child, held to a higher standard under UK data protection law and the ICO's Children's code than most of the data a bank holds.

DfE guidance and the updated Keeping Children Safe in Education framework both push schools towards caution about what is entered into AI tools, and towards not allowing pupils' original work to be used to train models without permission. A warning placed on a teacher is not a safeguard. It moves the risk to the person with the least time to carry it. The alternative is to build so the work never leaves the building, which removes the decision instead of delegating it.

What does good AI-drafted feedback actually look like?

It is specific, it is actionable, and it sounds like the person who teaches the class. Generic praise helps nobody, and a comment written in a register the pupil has never heard from their teacher is worse than no comment at all, because it tells the child the feedback was not really for them.

• It cites the pupil's own work, not a template. A comment that could be pasted onto any book in the pile is not feedback.
• It gives one next step, not five. A child who is told everything is told nothing.
• It is written for the reader. Reading age, vocabulary and sentence length matter more in feedback than almost anywhere else, because the pupil who most needs the comment is often the one least able to decode it.
• It is editable before it is sent. If the teacher cannot change it, the teacher does not own it.
• It leaves a record of what the teacher changed, so the school can see where the drafting is weak.

What we are building, and what we are not claiming

Marking and feedback is the module of Aevum Teacher Manager furthest along: the model is trained, on British national infrastructure, through the UK's AI Research Resource. It drafts against the school's own mark scheme, structured as what went well and what would make it better, and every piece of feedback goes to a teacher to review, edit and approve before a child ever sees it. It runs on the school's own hardware, so the work stays in the building.

What we are not doing is quoting you a figure for hours saved. We have not measured that in a real school year, and until we have, the honest status is that it is designed to do this and is being tested. It is in development. When there is a number, it will come with the protocol that produced it.

Own the model. Keep the data. Inspect the reasoning.