On September 2, 2026, New York City took student-facing generative AI out of every grade below ninth. The moratorium announced by Mayor Zohran Kwame Mamdani and Chancellor Kamar H. Samuels reaches about 600,000 children, bars companion chatbots at every age, caps daily screen time between grades 3 and 8, and leaves a pilot of five vetted tools open to no more than 50,000 high schoolers. It takes effect this school year.
Set aside whether you think that is brave or reckless. Look at what it took. Somebody held a mandate wide enough to stop a technology across a system that size, and the same announcement named a Technology in Schools Coalition to watch what happens and report back. A decision, and a date to look at it again.
Now ask what your school would do if a question of that shape landed on Thursday. Not what you would decide. Where.
Most schools have no answer. They have a data protection lead with a view, a deputy who came back from a conference full of it, a network manager who has read the terms and conditions nobody else has, and a head who gets asked about all of it in corridors. Each of those people knows something the others need. None can settle anything alone. And nothing in the calendar puts them in the same hour.
So the honest answer to who should sit on a school AI steering group is not a list of departments. It is five seats, drawn from the decisions the school owes an answer to: the person who can say yes, the person who carries the consequence when it goes wrong, the person who does the work, the person who will be asked about it in public, and the person who can say what is technically true. Seat those five and you have a group that can settle something. Seat the org chart and you have an audience.
Why a group picked from the org chart cannot decide anything
A group assembled to represent every department can describe a decision beautifully and make none.
This is not a failure of goodwill. It follows from how the seats were chosen. Representation asks who ought to be in the room; authority asks what the room is for. Pick by representation and you get eleven people, each carrying a brief to protect, none carrying permission to commit the school. The meeting fills. The agenda does not empty.
The sector's own numbers show the gap between having words about AI and being able to act on them. CoSN's U.S. State of EdTech 2026, drawn from 607 district responses across 44 states between January 14 and March 1, 2026, found 79 percent of districts with AI guidelines in place, up from 57 percent a year earlier. Only 38 percent had written an AI-specific policy, and 34 percent had addressed data privacy. The same survey names departmental silos as one of the standing obstacles to planning technology at all.
Read those figures together and the shortage is not guidance. Four districts in five have guidance. The shortage is a place where guidance turns into a decision with a name on it.
Guidance is a document. A decision has a name and a date on it.
The Room Test
Before a new AI group meets twice, run it through a single filter.
The Room Test is three questions I ask of any school AI steering group before it meets a second time: Which of the decisions this school owes an answer to can be settled in this room today, without anyone leaving to ask permission? When this group says yes, does anything change on Monday? And when one of its decisions turns out to be wrong, who in this room hears about it first, and on what date does the group look at it again? A group that can answer all three is steering. A group that cannot is a reading group with a standing invitation.
The second question does most of the work, because it is the one nobody can answer with an intention. If the group agrees that staff may use a particular tool for a particular job and a teacher's Monday looks exactly as it did before, the group did not decide; it expressed a preference. Something has to move: a line in the staff handbook, an account switched on, a sentence in a letter home.
This is offered as a way of seating a group, not as a validated instrument. It works best said out loud at the first meeting, while the membership is still easy to change.
The five seats, and what each one is for
Each seat exists because a specific kind of decision dies without it.
The person who can say yes. Someone in the room must be able to commit the school without going away to ask. In a single school that is usually the head or a deputy carrying delegated authority in writing; in a trust or district it may be an executive leader whose delegation is recorded rather than assumed. Diagnostic question: if this group agreed something this morning, whose signature would it still need this afternoon? If the answer is a name, invite the name.
The person who carries the consequence. The safeguarding lead, the data protection lead, whoever answers the phone when something has gone wrong. They are not there to slow the group down. They are there because a decision made without them is a decision someone else will have to defend, and the person who defends it should have helped make it. This is the seat that connects the group to who is accountable when an AI tool gets it wrong.
The person who does the work. One practitioner who will actually use the thing, not a representative of practitioners. A representative brings the staff room's mood. A user brings the sentence that ends an argument: that will not work on a Tuesday, because Tuesday is when we do reports. Rotate this seat by term if you like, but never leave it empty, and never let it be filled by whoever is most enthusiastic about AI. Enthusiasm is a useful trait and a poor sample.
The person who will be asked about it in public. A governor, a trustee, a board link, or whoever stands in front of parents and inspectors. Their job in the room is to ask the question they will later be asked: how would we explain this to somebody who thinks it is a bad idea? A decision that cannot survive that question in a quiet meeting will not survive it in a loud one.
The person who can say what is technically true. Network manager, data lead, whoever knows which account a tool actually runs under and what the contract says. Without this seat a group will spend an hour deciding something the systems already do, or forbidding something that was never possible. This is the seat that makes what a school should require before approving an AI tool a check rather than a hope.
Five seats, five people. If two of them are the same person, you have three seats and should say so rather than pretend.
Who should not have a standing seat
Two people belong in the room sometimes and on the membership list never.
The first is the supplier. Invite them to present, question them hard, then let them leave before the decision. A vendor in the room changes what gets said about the vendor, and the group's value is that it can say the unflattering thing.
The second is the enthusiast who has no other claim on a seat. Every school has someone who found AI early and has been generous with what they learned, and that person is genuinely valuable. But enthusiasm is not a seat. If they hold one of the five, seat them. If they do not, give them the work rather than the chair, and be plain about why.
What I Tell Leadership Teams
The most useful question I ask a new AI group is not about membership at all. It is: what has this group decided since it started meeting?
Across the leadership teams I work with, the AI group that achieves least is very often the best attended. That is not a coincidence. A large group is comfortable. Nothing is owned, everyone contributes, the minutes read well, and the next meeting inherits the same agenda with a later date at the top. I have sat in these rooms from three sides: as the classroom teacher whose week a decision changed, as the assistant head who had to make it stick, and later as the person responsible for digital strategy in a further education college, which mostly meant being handed the chair.
From the chair, the pattern is easy to see and awkward to say. A group with no authority will find work it can do, and the work it can do is discussion. So it discusses. Six months on it knows a great deal and has changed nothing, and because everybody has been busy, nobody can quite explain what went wrong.
The fix is unglamorous. Shrink the group until every person in it is there because a decision needs them, then give it a list of decisions with dates.
Your terms of reference, in five lines
A school AI steering group needs terms of reference short enough to read aloud.
- What this group decides. Name them: approved tools, staff permissions, what students may use and when, what goes to parents, what goes to the board.
- What it does not decide, and who does. Curriculum sits with curriculum leaders. Individual safeguarding cases sit with the safeguarding lead. Write the boundary down before the first hard case tests it.
- Who must be present for a decision to stand. Name the seats, not the people. A decision taken without the person who carries the consequence is provisional.
- Where decisions are written down, and who may read them. One list, one owner, visible to staff. A decision nobody can find is a decision that will be made again.
- When each decision comes back. Every entry carries a review date. This is the line most groups skip and the one that makes the rest worth writing.
That fifth line is not administrative tidiness. It is the whole argument for a standing group rather than a working party, and the ground keeps proving it. England's Department for Education published generative AI product safety standards for education in January 2025 and updated them on January 19, 2026 to take in cognitive development, emotional and social development, mental health and manipulation. A school that approved a tool against the original standards approved it against a shorter list than the one now in print. Nobody did anything wrong. The floor moved.
It is moving in statute too. Writing for the National Association of State Boards of Education in February 2026, Winona Hao recorded 34 states with AI guidance by the end of 2025, Louisiana standing up an AI committee in October 2025, North Carolina creating an AI Leadership Council, and Ohio becoming the first state to require every district to adopt a formal AI use policy by July 1, 2026. Hao's observation is the one to keep: states that adopted guidance early have come to treat it as a living document that must evolve. A living document needs somebody whose job it is to notice when it has stopped being true.
When a steering group is the wrong answer
Sometimes the honest move is not to form one.
A small primary school with nine staff does not need a committee; it needs twenty minutes at the end of an existing meeting, with the same five questions and the head writing the answer down. Creating a group there adds a calendar entry and subtracts nothing.
A school facing one unresolved question rather than a stream of them does not need a standing body either. It needs a working party with a deadline and a duty to disband. Standing groups accumulate; temporary ones deliver.
And there is a use of committees that leaders should recognize in themselves. Forming a group is one of the most respectable ways to postpone a decision. It looks like progress, it distributes the discomfort, and it buys two months. If you are convening a group because the decision is genuinely shared, convene it. If you are convening it because you would rather not be the one to say no to a supplier, or to the staff, then what you owe is the answer, not the agenda. That is a judgment about ownership before it is a judgment about structure, which is the argument in who should own AI strategy.
Where to start this term
Do not start with the membership list. Start with the decisions.
Write down every AI decision your school owes an answer to before the end of this term. Most leadership teams find between six and ten: which tools staff may use, which students may use and at what age, what happens to a student accused of using AI, what parents are told and when, who signs off a new request, and what you would do if a tool you already approved changed under you. Put a name and a date beside each one.
Then seat the room from that list. If a decision on it cannot be settled by the five people you have named, you are missing a seat. If a seat you have filled is not needed by anything on the list, you have a guest rather than a member.
The group that comes out of that exercise is usually smaller than the one the school was about to create, and it meets less often than anyone expected. It also decides things, which is the only feature that matters. If the list itself is the hard part, how to create an AI strategy people actually use is the piece of work that produces it.
If your leadership team is trying to work out who should be in the room and what that room is allowed to settle, this is the kind of work I support through AI strategy sessions for schools and trusts. You can also get one short piece of thinking on leading through AI each week through the newsletter.
Sources and further reading
- Mayor Mamdani and Chancellor Samuels Put Students First with Nation's Broadest Generative AI Moratorium in Schools, Office of the Mayor, City of New York, September 2, 2026.
- U.S. State of EdTech 2026, CoSN, May 2026, from 607 district responses collected January 14 to March 1, 2026.
- States Take Next Steps on Governing AI Use in Schools, Winona Hao, National Association of State Boards of Education, February 2026.
- Generative AI: product safety standards, Department for Education, published January 22, 2025, updated January 19, 2026.
Dan Fitzpatrick is the founder of The AI Educator, a Forbes contributor and the author of four bestselling books on AI in education. He works with school and organizational leaders on AI strategy, readiness and leadership development. More about Dan.
