A teacher uses AI to adapt a text for a pupil who joined a GCSE class late. Another uses it to draft quiz questions, checks them, and throws half away. A third is told to use the trust's chosen tool for planning and spends longer repairing its output than writing the lesson would have taken.
All three count as “using AI”. A policy that merely names an approved product has almost nothing to say about the difference.
That is the problem underneath a new Education Policy Institute paper, produced with the National Education Union after a roundtable of sector experts, trust leaders and academics. It asks how schools can support AI use while protecting teacher judgement and the relationships on which teaching depends.
The numbers explain why this conversation is overdue. The DfE's 2024–25 technology survey estimated that about two fifths of teachers used AI. In a separate NEU survey from February 2026, roughly half of teachers said their school had a policy on AI use by staff and students. Different surveys, different samples, but a recognisable pattern: practice is moving faster than the rules around it.
The approved-tool trap
Schools have good reasons to control which services staff use. Personal data, safeguarding, security and procurement do not vanish because a generated worksheet looks helpful. But a list of approved tools only answers where a teacher may go. It does not answer what for, with which information, or how the result should be checked.
The EPI roundtable heard from trusts that favour structured policies, named AI leads and controlled trials. It also heard from leaders who put more weight on trust and teacher autonomy. That tension is real. An open-ended “use your judgement” can leave staff guessing about privacy. A rigid tool mandate can override the subject judgement that makes a lesson worth teaching.
The middle ground is to govern the task. Can a teacher use AI to generate six distractors for a low-stakes quiz? What must they check? Could they paste in pupil work? Who approves a tool that students will use directly? Those are decisions a colleague can apply tomorrow morning.
Training has to meet the subject
The paper says some trusts found general AI training missed staff who needed to see their own subject in it. One moved toward subject-specific sessions. “Check the output” means one thing for an English model paragraph, another for a science explanation, and something different again for feedback to a pupil with additional needs.
The DfE already has free training materials. The paper notes that in an earlier government survey only 12% of primary leaders and 16% of secondary leaders said they knew a lot about those resources. Publication is not the same as uptake. Staff need time to try the material on a lesson they already teach and talk through what they would keep, edit or reject.
The authors recommend wider promotion of the DfE materials and more subject-specific examples. They also argue against treating workload savings as the only possible benefit. A teacher might spend the same amount of time but produce a better adaptation for one learner. That is a different claim and should be judged differently.
Write a policy a teacher can use
Take one task from each department: a quiz, a source sheet, feedback, an adapted reading passage. For each, record what AI may help with, what information may be entered, what the teacher must verify and how the final decision stays with them. Then ask the department to try the rule on an actual example.
That exercise will expose more than a generic acceptable-use paragraph. It will show where the policy is too vague, where a tool is genuinely useful and where it quietly shifts work or responsibility back onto the teacher. A good AI policy should help a teacher make those distinctions before the next lesson, not after something has gone wrong.
Sources: Education Policy Institute and National Education Union, event summary paper, 22 September 2026; DfE, Technology in Schools Survey 2024 to 2025. The EPI paper records a roundtable discussion; its recommendations are not government policy.