In the first week of September 2026, New York City told roughly 600,000 pupils, two thirds of its roll, that they would not be using generative AI in school this year. Within days Los Angeles Unified, the second largest district in the United States, did the same for every student on a district-issued device. And the Royal Society's written evidence to the House of Commons Education Committee, drawing on a Teacher Tapp survey of more than 9,000 teachers in England conducted in March 2026, found that only 14% felt confident across all the dimensions of AI literacy and only 15% had whole-school guidance to work from.
Put those three things together and the temptation is obvious. If most staff are not ready and the biggest school systems in the world are pulling back, restriction looks like the safe move. Since the announcements, almost every school leader I have spoken to has asked me some version of the same question: should we?
Should schools ban AI?
Schools should not ban AI by default, and most should not ban it outright. But restricting pupils' use of it is defensible when three things are true at once: the thing you are protecting is specific, you cannot yet protect it any other way, and you can say what you would need to see before you lifted the restriction.
A restriction that meets those three tests is a decision. One that does not is a school drifting into prohibition because nobody chose anything else, and a drift cannot be defended to a parent, a governor or an inspector, who is more likely to ask how you make sure your use of AI serves children's best interests than to ask for your policy. The rest of this article is the practical version: the three positions actually open to you, the test to run before you take one, and the part almost nobody plans for, which is how to lift a restriction without the reversal costing you more than the ban did.
A ban is a decision, not a default
Restricting pupils' use of AI is a leadership decision with consequences, and it deserves the same rigor you would apply to introducing it. The uncomfortable truth is that most of the bans I come across in schools are not decisions at all. Nobody chose to restrict. The filter was never opened, the policy was never written, and the school drifted into prohibition because prohibition is what happens when no one decides.
The difference matters because a default cannot be defended. When a parent asks why their child cannot use at school a tool they use every evening at home, "we have not got round to it" is not an answer. Neither is "New York did it."
New York, in fact, has been here before. It blocked ChatGPT on school networks in January 2023 and became the first district to reverse such a ban, four months later, as Regina Ta and Darrell M. West recorded in their Brookings Institution analysis of banning versus integrating generative AI in August 2023.
What is different in 2026 is that the restriction has been made in the open. The city's announcement from Mayor Mamdani and Chancellor Samuels states its reason plainly — "Children need teachers and human connection in order to learn and grow" — and then does the work a defensible restriction has to do. It covers 2-K through eighth grade. It exempts teachers, who may keep using AI for planning and operational tasks. It carves out assistive technology for students with disabilities, support for multilingual learners and career readiness programs such as computer science. It runs five supervised high school pilots reaching up to 50,000 students. It gives every high schooler two 45-minute AI literacy modules a year covering what AI is and is not, bias, risk and ethics. And it hands a Technology in Schools Coalition of students, educators, parents, officials and experts the job of assessing the moratorium across the 2026-27 school year and publishing recommendations. You can disagree with the decision. You cannot say it was not made as one.
Los Angeles is the cautionary contrast. Its restriction on all student access to generative AI platforms was confirmed by email on September 2, 2026, and, as K-12 Dive reported, a Generative AI Ad Hoc Committee met the following day and will bring recommendations to the board by the end of the 2026-27 school year. The problem was the route the decision took. Senior administrators enacted it on their own authority, and board member Kelly Gonez told Education Week that she "even as a board member, was under the expectation that we were still operating under last year's policy."
Whatever the merits of the restriction, a board that learns of it afterwards is a governance failure, and it is the first lesson for anyone watching from a distance. Working out who has to be in the room before a decision of this size leaves it is its own piece of work, and who should sit on your school's AI steering group sets out the membership that prevents exactly this.
The three positions a school can take
There are three positions available, not two, and the mistake most leadership teams make is to argue about "ban or allow" when the real choice is between the second and the third.
Prohibition. Pupils may not use generative AI for schoolwork, indefinitely, with no review date. Described plainly, this is a hope that the question will go away. It will not.
A moratorium with an end. Pupils may not use it for a defined period while the school does named work: staff training, safeguarding checks, tool vetting, assessment redesign. There is a date on which the decision is revisited and a body responsible for revisiting it. Both New York and Los Angeles sit here, each with a named group reporting within the school year.
Staged permission. Pupils may use a short list of approved tools, for named purposes, from a named age or year group, and the list widens as evidence accumulates. This is where Los Angeles Unified stood until September: students aged 13 and over could use approved generative AI tools once they had completed mandatory digital citizenship instruction and reviewed the district's responsible use policy. Keeping such a list short and defensible is a discipline of its own, and what a school should require before approving an AI tool sets out the questions that do it.
A prohibition answers the question "may they?". A moratorium answers "not yet, until what?". Staged permission answers "who, for what, and how will we know?". Only the last two are leadership positions, because only the last two commit the school to doing something.
When is restricting AI the right call?
Restriction is the right call when the thing you are protecting is specific, the school cannot yet protect it any other way, and you can say when that will change.
The strongest case is with younger children. I have written before about the Handover Question, the one test I ask school leaders to put to every task children are set: if a machine did this for them, what would they lose? For a seven-year-old learning to compose a paragraph or hold a number in their head, the answer is the learning itself. Protecting that is not technophobia. It is knowing what the task is for. New York's decision to draw its line at the end of middle school, and to exempt assistive technology and multilingual learners, is a recognizable version of that logic.
The second case is safeguarding. New York's restriction names companion chatbots specifically, and the direction of UK policy points the same way. On June 15, 2026 the government set out plans to require AI "romantic companion" chatbots to enforce a minimum age of 18, with intimate functions restricted for under-18s across all chatbots; the University of Oxford's expert comment on the announcement notes that legislation was expected before Parliament by the end of 2026, with the first restrictions potentially taking effect in spring 2027. That is a proposal rather than law, so plan against the duties you already have. But a school that has not yet worked out which tools children are talking to, and what those tools say back, has a reason to pause.
The third case is capacity. The Royal Society's written evidence to the Education Committee found that 23% of teachers in England receive no AI support of any kind, and that teachers with both structured professional development and whole-school guidance are roughly three times more likely to be confident across all dimensions of AI literacy than those receiving neither. If your staff cannot yet supervise pupil use well, a moratorium that buys time to change that is defensible. A moratorium that buys time and then wastes it is not.
Restriction is the wrong call when the reason is a headline, when the reason is cheating, or when the reason is that the leadership team cannot agree. Bans do not stop cheating; they move it home, where you cannot see it and cannot teach through it, and they leave you handling suspicion with no evidence, which is the hardest version of what a school should do when a student is accused of using AI. The durable answer to integrity is not prohibition but design, which is the argument of what happens to assessment when AI can produce excellent work. And a ban adopted to postpone an argument is the worst kind, because the argument is still waiting when the ban is lifted.
The Restriction Test
The Restriction Test is three questions a leadership team must be able to answer before it restricts pupils' use of AI: What exactly are we protecting? What are we giving up to protect it? And what would we need to see before we lifted it? A restriction with three answers is a decision. A restriction without them is a mood.
I offer it as a way of thinking rather than a validated instrument, and its only real claim is that a leadership team which cannot fill in all three lines has not yet made a decision.
What exactly are we protecting? The weak answer is "children". The strong answer names the thing: independent composition and handwriting fluency in the early years; the habit of asking an adult rather than a chatbot; pupils' data, which the school has not yet confirmed is safe with the tools in question. If you cannot name it, you cannot tell when it is safe.
What are we giving up to protect it? The weak answer is "nothing". There is always something. Older pupils will leave without supervised practice in a skill their next school, college or employer will assume. The pupils whose families provide tools at home will get that practice anyway. Jean-Claude Brizard, chief executive of Digital Promise, put that objection bluntly to Education Week on September 2, 2026: "Kids of middle income, upper income parents are going to have access to technology. They're going to learn how to use AI. The poor kids in New York City will be the ones who will suffer [without a] sensible approach to leveraging technology."
The Royal Society found the same gradient in England, where 42% of teachers in the most deprived schools said AI literacy is not addressed in their teaching at all, against 33% in the most affluent. Write the loss down, so that the restriction is a trade you have chosen rather than a loss you have ignored.
What would we need to see before we lifted it? The weak answer is "we will review it". The strong answer is a date and a list: staff trained and guidance issued; a short approved-tools list agreed; assessment in the affected year groups redesigned so that process and live performance count, not just the product; safeguarding checks on any tool a pupil will talk to. This is the same discipline as the Exit Question I ask of every pilot. A pause with no exit is not a pause. It is a habit.
How do you lift a restriction without the reversal costing more than the ban?
You lift a restriction by showing that the conditions you named have changed, in the order you named them, rather than by announcing that the pause is over.
This is the half of the decision nobody plans, and New York's 2023 experience is the warning. The ban itself was defensible. Reversing it within four months, on the reasoning that the original block had been a reaction rather than a judgment, made both ends of it look unconsidered. The restriction cost the district little. The reversal cost it the argument.
Both districts have now bought themselves a school year and named a body to report. That is the easy part. The hard part arrives in June, when the report lands in a different political weather from the one the restriction was made in. Four moves, in this order, are what separate a lift from a climbdown.
Put the three answers back on the table and mark each one. Met, partly met, or not met. The Restriction Test is not only an entry test; it is the exit criteria you wrote for yourself. If you protected early composition, show what changed about how that is taught and assessed. If you paused for staff capacity, show the training and the guidance, both of them, because the Royal Society's finding is that neither works alone.
Lift narrowly first. A named year group, a named short list of tools, a named purpose. Staged permission is the landing place, not full permission. A lift that goes straight from nothing to everything tells staff that the restriction was never about the reasons you gave.
Say what you learned, including what you got wrong. A school that can say "we paused, here is what we did with the time, here is what we now allow and why" has converted a restriction into credibility. A school that quietly reopens the filter has taught everyone that its policies are weather.
Publish the lift where you published the ban. Parents and staff who heard the restriction from a letter must hear the lift from a letter, not from their children. What to put in the letter home about AI applies as much to lifting a restriction as to imposing one.
Underneath all four sits the test worth keeping: a restriction lifted because the condition was met is governance; a restriction lifted because the noise died down is weather. If you cannot tell which one you are doing, it is the second.
What makes the lift possible is the work done during the pause, and that work is mostly not about AI tools at all. It is knowing where the school actually stands, which is what the seven dimensions of AI readiness are for; and having decided what AI is for in your school, which is the job of an AI strategy people actually use. A leader newly in post who has inherited a restriction should treat the review date as a fixed point and work backward from it, which is roughly how the first 90 days of leading AI change sequences the rest.
What Good Looks Like
A restriction done well is short, written down, dated, and accompanied by the work the school will do while it is in force.
As a trustee, I have sat on the other side of this table, and the version of a restriction that a board can support has a consistent shape. It distinguishes pupils from staff: both New York and Los Angeles kept AI in teachers' hands while removing it from students', and the Royal Society's figures show why, because staff need the practice more than anyone. It names its end, or at least its review. It says what the school will do in the meantime, and who owns each piece. It reaches parents and governors before it is announced rather than after, because a policy families first meet in the press is a policy you will spend a term defending instead of implementing. And it keeps teaching about AI while restricting the use of it. A ban on use is not a ban on learning, and a school that confuses the two ends up with children who have never used AI at school and never understood it either.
Having trained educators in more than 30 countries, I have watched this cycle more than once: the ban, the quiet workaround, the reversal. The schools that come out of it well are not the ones that chose the right side. They are the ones that could explain their choice, and changed it when the evidence said to.
What should leaders do this term?
Write down the three answers, because a leadership team that can produce them on one page has already done most of the work.
Then separate staff from pupils, and make sure any restriction on the second is not accidentally a restriction on the first. Put the review date in the diary and name who owns it. Tell governors and parents before you announce, and give them the three answers rather than the headline. Decide now, while nobody is under pressure, what evidence would justify lifting it. And if you choose a moratorium, spend it: the Royal Society's finding that training plus whole-school guidance roughly triples staff confidence is the clearest description I have seen of what the time is for.
Above all, notice that whichever position you take, you are still left with the strategy question. A restriction only prevents; it does not tell anyone what AI is for in your school. That answer still has to be written, and the schools that use a pause to write it will be the ones that lift the pause well.
The next step
If your leadership team or board is working through this decision now, this is the kind of question I help schools and trusts settle through AI strategy sessions for schools: not which side to take, but how to take a side you can defend, and how to know when to change it. If you would rather read than book anything, I write a newsletter for school and trust leaders on the decisions this kind of work turns on.
Sources and further reading
- Office of the Mayor of New York City, "Mayor Mamdani and Chancellor Samuels Put Students First with Nation's Broadest Generative AI Moratorium in Schools," September 2, 2026. nyc.gov
- Klein, A., "New York Restricts AI in Elementary and Middle Schools. Will Other Districts Follow?," Education Week, September 2, 2026 (updated September 3, 2026). edweek.org
- "LAUSD Adopts Moratorium on Student AI Use, Another Move to Limit Classroom Tech," Education Week, September 2026. edweek.org
- "LAUSD restricts all students from using AI tools," K-12 Dive, September 2026. k12dive.com
- Written evidence from the Royal Society to the House of Commons Education Committee (AIE0215), including a Teacher Tapp survey of more than 9,000 teachers in England, fieldwork March 2026. committees.parliament.uk
- Ta, R. and West, D. M., "Should schools ban or integrate generative AI in the classroom?," Brookings Institution, August 7, 2023. brookings.edu
- "Expert Comment: Social media and AI chatbot limits for children — what does the evidence support?," University of Oxford, June 16, 2026. ox.ac.uk
Dan Fitzpatrick is the founder of The AI Educator, a school trustee, a Forbes contributor and the author of bestselling books on AI in education, and has trained more than 150,000 educators across 30 countries. More about Dan.
