My daughter is six. At dinner one evening I mentioned that I had spent the day writing about AI, and she sat up and said: “Say no to AI.”
I asked her where that came from. Older kids at school, she said.
Were they right?
In this post I will:
Set out the real reasons two enormous school systems have switched AI off.
Show what happened the last two times we tried prohibition in education.
Give you five things to do instead.
What New York and Los Angeles decided
On 2 September 2026, New York City imposed a one-year moratorium on student-facing generative AI for every child up to the age of about fourteen. That is nearly 600,000 children, or two-thirds of all children in the public school system. Companion chatbots are prohibited for all ages. A maximum of 50,000 secondary students, about 5% of enrolment, get five approved tools, each capped at a certain number of minutes a week.
Los Angeles Unified had already gone further. Generative AI has been switched off for every student since the start of this school year, and at a board committee that same week its chair, Kelly Gonez, summed it up:
“Currently, there is no access to generative AI for any student. I think that’s probably the headline.”
Federal figures put its enrolment at 522,080 in 2025-26.
Between the two jurisdictions, this is well over a million children.
Why a school system would do this
The version of this case that circulates online is a caricature about cheating. The real reasons are better:
There is a genuine legal issue. A district cannot lawfully hand a child’s data to a vendor it has not assessed, and every tool needs an assessment, a contract, an age check and a retention policy. There are hundreds of these AI products and they change monthly. Switching off by default is often the only lawful position available to an organisation that has run out of money to fund lawyers.
There is a mounting body of education research outlining the dangers of AI overuse in the classroom. A field experiment with nearly a thousand high school mathematics students in Türkiye, published in PNAS in June 2025, gave one group access to GPT-4 while they practised. Their grades improved substantially. Then the researchers took it away, and those students performed worse than the group that never had it at all, by 17%. They had used AI as a crutch, and when the crutch went the walking had not been learned.
Some AI tools are more equal. A ban is the only intervention a school can apply to every child at once. Everything else depends on which teacher you get, how confident your parents are, and whether your household can afford the pro version of the tool. Telling a head to run a nuanced tool-by-tool policy instead is telling them to widen a gap they are employed to close.
The chancellor of New York City Public Schools, Kamar Samuels laid out their justification clearly:
“We’re putting guardrails in place to protect the human connection, curiosity and creativity that help children grow. We’re standing firmly in our belief that innovation does not mean more technology.”
I disagree with the conclusion. However, I honestly see where Chancellor Samuels is coming from, and why he arrived where he did.
The last two times we tried this
Abstinence-only sex education is the closest fit, because it is education policy and because somebody tested it properly. The United States government commissioned a randomised controlled trial of four programmes, following more than 2,000 teenagers. Mathematica reported in 2007 that
“Programs had no effect on the sexual abstinence of youth”,
and that
“Youth in these programs were no more likely to have unprotected sex.”
A decade later a review in the Journal of Adolescent Health found the same, and noted that the funding continued regardless. Indeed, this research found that these programs:
“…inherently withhold information about human sexuality and may provide medically inaccurate and stigmatizing information.”
School phone bans also need considering, because the prohibition now being extended to AI has already been tested on its own terms. A national American working paper, not yet peer reviewed, used GPS data and lockable pouches and found phone use substantially reduced and average test score effects consistently close to zero. They also found:
“…little evidence of effects on school attendance and no clear evidence of improvements in self-reported classroom attention or perceived online bullying.”
The problem is that any perceived benefit lives in compliance, and these policies only ever covered the network. Los Angeles has already said that its moratorium reaches district-issued devices and that the board cannot regulate the phone in a child’s pocket.
The PNAS study in Türkiye has a second half that is rarely quoted. The same researchers ran a version called GPT Tutor: the identical model, with guardrails built into how it answered. The crutch effect largely disappeared. What fixed it was how the thing was built.
Who a ban lands on
Pew surveyed 1,391 American teenagers in late 2024 and found that 26% had used ChatGPT for schoolwork, double the previous year. By race: 31% of Black teenagers, 31% of Hispanic teenagers, 22% of White teenagers. A year earlier those three figures sat level with one another.
A school ban removes the supervised, teacher-visible version of a tool. It does not remove the tool. The child with a laptop at home and a parent who knows about fabricated references, context rot, and simulated empathy, carries on unaffected. The child whose only supervised access was a school device loses the supervision and keeps the phone.
My book goes underneath all of this. Slow AI is about the question this post is a corner of: knowing when to use these tools, and when to leave them alone. It was a #1 Amazon bestseller in Education.
The one restriction I support
Companion chatbots should not be available to children, and New York City is right to prohibit them for all ages.
NORC surveyed 1,060 American teenagers for Common Sense Media in 2025. 72% had used AI companions, and a third of those users had talked to one instead of a person about something serious. The report’s own conclusion is that
“Given the current state of AI platforms, no one younger than 18 should use AI companions.”
The intervention that has actually changed what children experience came from the company, not the school. Character.AI removed open-ended chat for under-18s, calling it one of the most significant safety decisions it has made. That is a design change, and it reached every child on every device, including all the ones a school district cannot see.
It is the same finding as GPT Tutor. Change how the thing behaves and the harm moves. Remove access at the school gate and the harm walks home.
Five things to do instead of AI prohibition
1. Ask the children, and write down what they say. You do not need a working group. A parent can ask at dinner what their child thinks AI is for, and a teacher can spend ten minutes of one lesson asking a class what they already use and what they worry about. New York convened its student coalition after the ban. My daughter formed a position without anybody asking her anything.
2. Teach these three questions. Not how to prompt. When is this system guessing, who profits from me believing it, and what is it doing to my own thinking while I use it. Those three work in a classroom, a staff meeting, and a kitchen.
3. Have the conversations at home. “What do you know about AI?” “Are you using it in your classrooms?” “If it tells you something important, who else can you ask?” No school policy reached my kitchen table, and none will reach yours.
4. Try the things yourself before you decide about them. You cannot judge the limits of a tool you have never opened, and most of the adults legislating about AI have never used it regularly.
5. Ask what the design change would be. Before a restriction, ask what could be altered in how the thing behaves, and who you would have to ask. Restriction is what you reach for when nobody will change the product.
California legislated the alternative in 2024, and its curriculum commission must consider incorporating AI literacy whenever the frameworks are next revised. Two of the three were last adopted in 2016.
Prohibition can be implemented in a fortnight and educational change can take a decade (or longer!), so of course one of them keeps arriving first.
What I have not worked out
New York and Los Angeles are led by people who read the same material I do and reached a different conclusion. A one-year moratorium with an evaluation attached is a study that I am really looking forward to analysing.
Life does not work in solid lines, though. It works by osmosis, across semi-permeable membranes, which is how a six-year-old ends up holding a policy position she got from children four years older than her.
My daughter told me to say no to AI. I have just spent almost two thousand words arguing against saying no. I have not yet worked out what to say back to her, and I am not sure the answer is in any of the studies above.
Tell me in the comments what you would say, particularly if you disagree with me. I would like to hear from teachers and administrators who have watched one of these policies arrive, and from parents who have had the conversation already.
Go slow.




The comparison with abstinence-only education and school phone bans rests on a category error. Abstinence-only education failed for reasons specific to that policy: it withheld practical information about behavior adolescents would continue to encounter and left them poorly equipped to manage real-world health risks. School phone bans address a different mechanism. Smartphones provide continuous access to socially reinforcing, dopamine-driven reward loops that encourage compulsive checking, fragment attention, and compete with classroom instruction. Their value also extends beyond any immediate change in test scores.
Generative AI presents another distinct policy problem involving cognitive development, academic integrity, student data, tool design, and the delegation of intellectual work. Placing all three policies under the label of “prohibition” obscures the mechanisms that determine whether each intervention works. This seems to call for a “mu” response because the comparison itself lacks a sound premise. Your strongest evidence comes from the AI research comparing unrestricted GPT access with the guarded tutoring model, which directly supports policies tailored to specific uses, ages, safeguards, and learning goals.
Pushing the discussion home assumes that parents are broadly informed about the generative AI ecosystem—something that I suspect is generally not so. Although the LA and NYC district leaders *might* have read the same things about AI as you have, they are additionally subject to the political pressure of those same underinformed, and currently tech-panicking, parents.