Give every task a teacher might hand to AI one of three answers: encourage it, allow it with a named check, or not yet. Put the answers on one page, name the person who owns them, and review the page every half-term. Most schools have not done this. In a Gallup and Walton Family Foundation survey of 2,069 U.S. public school teachers, fielded from February 9 to March 2, 2026, only 18 percent said they had received any formal guidance from administrators on how AI should be used. About a third (34 percent) had no guidance on any of the ten tasks the survey asked about, and 48 percent had only informal guidance, which means a colleague's opinion in the staff room.
A school can have an approved AI policy and still sit in that 82 percent. That is the gap this article is about.
Why do most teachers have no AI guidance?
Most teachers have no AI guidance because schools wrote a policy and stopped. A policy tells a teacher what the school believes. Guidance tells a teacher what to do at nine o'clock on Sunday night with thirty essays and a tool open in another tab. The first is a governance document, approved once. The second has to be usable by one tired person, alone, in under a minute.
The evidence on the surrounding support points the same way. Education Week reported in May 2026, drawing on its EdWeek Research Center surveys, that the share of teachers with no AI training fell from almost 60 percent in October 2024 to 42 percent in winter 2026, while only 9 percent reported ongoing training. It described the training that does exist as uneven at best in frequency, type and depth, and it noted that teacher reluctance partly stems from a lack of clear leadership guidance. Training moves skill into a person. It cannot tell that person what the school has decided.
One caution on the sources. Both are American, and the Gallup sample is U.S. public school teachers. I have not tested whether the same figures hold in England, so treat the percentages as a picture of one system. The pattern will be familiar to any UK leader who has asked a room of teachers what they may use AI for and watched the room go quiet.
Where is the guidance gap widest?
The gap is widest on the two tasks where a mistake does the most damage to a child. Gallup found that 69 percent of teachers get no guidance on one-on-one instruction or tutoring, and 58 percent get none on grading and student feedback. Compare that with the easy wins: among teachers who do get guidance, encouragement is most common for modifying student materials (58 percent), making worksheets (54 percent) and preparing to teach (53 percent).
Look at grading more closely. Among teachers who receive guidance on grading, 10 percent say they are discouraged from using AI for it, the highest figure for any task. Another 54 percent say the guidance neither encourages nor discourages use. So where schools have said something about grading, more than half said nothing that tells a teacher which way to go.
Tutoring has a different problem: the evidence itself is thin. In June 2026, Tes reported that the Department for Science, Innovation and Technology described AI tutoring tools as "limited in quantity, scope and evidence base" in the contract notice for its AI tutor pilot, with school trials planned from the autumn term of 2026. A leader who wants to write tutoring guidance cannot lean on a body of evidence. That is a reason to write a narrower stance, not to write nothing.
Here is my interpretation. Silence is not neutral. A teacher who hears nothing about grading will do one of two things: the confident teacher will assume permission, and the careful teacher will assume prohibition. You then have two standards in one building, and neither was chosen. Gallup's finding on written guidance supports acting: 69 percent of teachers with formal guidance on making student materials said it encouraged AI use, against 51 percent of those with informal guidance. Writing a thing down changes what people feel able to do.
What does good AI guidance look like?
Good AI guidance passes the Sunday Night Test. Here is the definition I use.
The Sunday Night Test is three questions I ask of any AI guidance a school has written: Could a teacher alone with a pile of marking find the answer for this task in under a minute? Does the answer name the check that has to happen before anything leaves the room? And does it say who to ask when the case does not fit? Guidance that passes all three is something a teacher can use. Guidance that fails any of them is a statement of values, and teachers will fill the gap with their own judgment, privately.
This is my suggested way of testing guidance, not a validated instrument. It exists because most AI documents fail the first question before they reach the other two. A twelve-page policy does not answer "may I use AI to write feedback on these essays?" in under a minute.
The format that passes is a table with one row per task. Each row carries a stance, a check and an owner. The stances are only three: encourage, allow with a named check, and not yet. Below is a starting point. These are my suggested defaults, offered as a hypothetical first draft for a school that has not started. Adjust them to your local law, your data protection officer's advice and the tools your school has approved.
| Task | Suggested stance | The named check | Who to ask |
|---|---|---|---|
| Lesson planning and preparation | Encourage | Read every output; check facts against your curriculum | Subject lead |
| Adapting materials (reading level, worksheets) | Encourage | Check for errors and for tone with your actual class in mind | Subject lead |
| Parent and carer emails | Allow with a check | No names or personal details go into the tool; you read it and send it as yourself | Line manager |
| Marking and feedback on student work | Allow with a named check | You decide the grade; you read every comment before a student sees it; only tools on the approved list | AI lead |
| Reports and support plans | Allow for wording only | No personal data in any tool that is not approved; the teacher owns every judgment in the text | AI lead and SENCO |
| One-on-one AI tutoring for students | Not yet | Pilot only: a named owner, a small group, a review date | Principal or headteacher |
| Decisions about an individual child (grades, discipline, safeguarding) | Not with AI | The decision stays with a named person | Designated safeguarding lead |
Three of those rows lean on other work. The approved-tools column depends on a list you can defend, and approving AI tools for schools covers how to build one. The reports row depends on a position about whether AI should write school reports and, for support plans, on what a SENCO should and should not hand to AI. The last row is the boundary I argue for in which decisions should AI never make.
The Mistake I See Most Often
The mistake I see most often is a finished policy and unstarted guidance. Across the leadership teams I work with, the policy is usually good. Governors have approved it, and it sits in the shared drive. Then a teacher asks a colleague whether she may use AI to draft feedback, the colleague says she thinks so, and that answer becomes the school's actual position.
I spent years as a teacher and an assistant headteacher before I moved into advisory work, and the questions that matter arrive in the staff room, not in the policy folder. They are small and urgent and they come from people who want to do the right thing. A good policy cannot answer them. A good table can, which is why I would write the table before I revised the policy a second time. If your policy is the next job on the list, a short school AI policy is a sensible place to start, and the table slots in beneath it.
How do you write the first version this half-term?
You can write a usable first version in about a week, with three teachers and one afternoon of your own time. Work in this order.
- List the tasks teachers actually did with AI in the last two weeks, not the tasks you hope they did. Ask in a staff meeting and promise that nobody will be told off for answering. If people hold back, you have learned something about how safe it feels to be honest.
- Give each task one of the three stances. Where you cannot name a check that a teacher could carry out, the stance is "not yet", and that is a legitimate answer.
- Name an owner for every row. A row without a name is a wish.
- Hand the draft to a teacher who did not help write it and ask her to find the answer for marking in under a minute. If she cannot, the table has failed the Sunday Night Test, and the fault is in the table.
- Put a review date on the page. Half-term is a natural rhythm, because tools change faster than governing bodies meet.
Two objections are worth answering. The first is that guidance dates quickly. It does, which is why it lives in a table with a review date and an owner rather than in a policy that needs a board to change it. The second is that a "not yet" on tutoring or grading looks timid. I think a bounded yes with a real check beats silence, and a clear "not yet" beats a vague "be careful". What does not work is saying nothing and hoping the careful teachers and the confident teachers reach the same conclusion.
Once the table exists, the training question gets easier too, because you can teach to it. Skills for staff beyond basic AI training covers what comes next, and what to do when staff get AI wrong covers the other side of the same coin: a school can only respond fairly to a mistake if it said, beforehand, what was expected.
If your leadership team is working through these questions, this is the kind of work I support through AI training for schools and advisory sessions with senior teams.
Sources and further reading
- Gallup and the Walton Family Foundation, "Most Teachers Receive No Formal Guidance on AI Use," Gallup, May 26, 2026 (survey of 2,069 U.S. public K-12 teachers drawn from the RAND American Teacher Panel, February 9 to March 2, 2026). news.gallup.com
- Elizabeth Heubeck, "More Schools Are Providing AI Training for Teachers. Is It Any Good?," Education Week, May 18, 2026. edweek.org
- Cerys Turner, "'Limited' evidence for AI tutoring tools, government says," Tes, June 17, 2026 (updated June 18, 2026). tes.com
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and a bestselling author on AI in education.
