30 October 2026
Ask for the override number

4 min read
A board cannot judge an AI tool by reading its policy. It can ask how often a teacher changed what the tool drafted, and the school should be able to answer from its own records.
A school board is now expected to say whether the AI used in its school is safe, ethical and accountable. A Schools Week opinion piece published on 22 October by Samira Sadeghi of the Confederation of School Trusts makes the case that boards do not need to be AI experts, but do need to be able to answer that question.
I agree with her. The difficulty is what a governor is handed to answer it with. A policy says what should happen. A supplier's assurance says what the supplier believes. Neither tells the board what happened in a classroom on a Tuesday.
One number a board can ask for
How often did a teacher change what the AI drafted before it reached a pupil or a parent?
If the answer is never, one of two things is true. Either the tool is perfect, or nobody is checking. The board cannot tell which from the number alone, and that is useful to know. If the answer is every time, the tool is not saving anyone time, and a teacher is doing the work twice. Somewhere between those two is what healthy review looks like.
The figure matters most as it moves. If it drops sharply after a software update, something changed. Either the tool improved, or people stopped looking.
Why this number and not another
Accuracy scores measure a model on a test set somebody else chose. The override number measures how your own teachers used the tool in your own school. It is also cheap to produce. The draft exists, the final version exists, and the difference between them is a record. No governor has to understand how a model works to read it.
Whose records are they?
This is where it usually goes wrong. If the only copy of the drafts sits on the supplier's servers, the board has to ask the supplier for the number. A school that must ask permission to see how its own staff used a tool is not in control of that tool.
So the number should come from the school's own records, without a request to anyone. I hold myself to that as a design requirement. I build AI for schools that runs on the school's own hardware, so the draft and the final version stay in the building and the log belongs to the school. Aevum Foundry AI's products are in development. I am describing a requirement we work to, not announcing a feature.
What the number cannot tell you
It does not say why. A teacher may change a draft for tone and not for error. A tired teacher may accept anything. So treat the figure as the start of a question and not as a verdict. Pair it with a sample: a board member reads ten drafts alongside the edits that were made to them.
We learned this from the other side. When we trained a model to adapt material better for SEND, it got worse at noticing a child in trouble. We only found that because we looked at what the model actually did, not only at how it scored. We published it: the regression we found in our own model. A school deserves a way to do the same check for itself.
Why it matters most for SEND
Updated government-backed guidance, reported by Schools Week, says AI should not write contributions to an EHC plan without substantial professional review. That is a sensible rule and it has a gap. Substantial review is something a school can only show if it recorded it. A teacher who reviewed carefully and a teacher who clicked accept look identical unless the record shows what changed.
The drafts that matter most in SEND are support plans, adapted materials and letters home about a child's needs. Those are exactly where a rubber stamp does the most harm, and where a board most needs evidence.
A question for schools
If your school uses AI to draft anything today, can you tell how often a teacher changes it? I taught for twenty years, and in most staffrooms I knew the honest answer would have been to ask around.
I am talking to SENCOs, heads and governors about how they would check this. If that is you, message me and tell me how it works in your school.
Last reviewed 30 October 2026.
Changelog
30 October 2026: article added.