A Ban Is Not a Plan. Neither Is Access.
In one week, New York and Los Angeles switched student AI off, the UAE extended compulsory AI to every private school, and the OECD published the number that says both are aiming at the wrong variable.
What this is: a parent's-eye read of one very strange week in AI-and-schools policy, and an argument that the two loudest answers on offer, banning AI and mandating it, share the same blind spot. We build a kids' AI maker, so we have a stake in this. We're saying so up front, and we're keeping our own product to one paragraph near the end.
The week
On Tuesday, September 1, New York City announced a one-year moratorium on student-facing generative AI for grades 2-K through 8, roughly 600,000 students, along with screen-time recommendations of 30 minutes a day for grades 3 to 5 and 45 minutes for grades 6 to 8. High schoolers get a couple of critical-thinking modules a year and a few restricted pilots. Teachers keep AI for planning.
On Wednesday, September 2, Los Angeles Unified, the second-largest district in the country, switched on software that blocks all generative AI on district-issued student devices. That covers about 378,000 students in every grade. The district announced it at the first meeting of a new AI committee, to the visible surprise of board members and parents in the room. The block holds while the committee reviews "instructional use and safeguards," with recommendations due by the end of the school year. So, effectively, a year, like New York.
The same week, the United Arab Emirates extended its compulsory AI curriculum, which has run in every public school from kindergarten through grade 12 since the 2025-26 year, to private schools as well, and put roughly 22,000 teachers through AI training ahead of the new term. Five-year-olds in Abu Dhabi are getting a weekly AI class. Eight-year-olds in Brooklyn and Boyle Heights are getting a block page.
As one K-12 advocate put it in the middle of all this: a ban is not a plan. She's right. But neither is access.
The number that undercuts both
Also this year, the OECD published its Digital Education Outlook 2026, the closest thing the field has to a consensus read of the evidence. Its finding on general-purpose generative AI is the one every school board should have taped to the wall:
Students with access to general-purpose generative AI produce better work. Then, when the tool is taken away, the advantage disappears, and sometimes reverses.
The report leans on field experiments, including one in Turkish high schools where AI access raised practice scores by roughly 127 percent, and the same students then scored about 17 percent worse than their peers on the unassisted exam. The OECD's own framing is blunt: outsourcing a task to generative AI without pedagogical design "simply enhances performance with no real learning gains." Where the AI is deliberately designed into how the learning happens, the report finds the opposite: more durable learning and stronger critical thinking, creativity and collaboration.
Read that against the week's headlines. New York and Los Angeles are turning access off. The UAE is turning it on for everyone. The OECD is saying access was never the variable. Whether the kid learns anything depends on what the AI is for in the room, not whether it is in the room.
Ban, mandate, design
It helps to name the three positions plainly.
Ban. Remove student-facing AI for a year, study it, come back with guardrails. The honest case for this is that nobody has a playbook and the downside cases are real. The honest problem is that the kids do not stop. They go home and open ChatGPT. The district has not removed AI from their lives; it has removed the one adult-supervised setting where someone could have shown them how to use it well. And it has done so unevenly: the families with a parent who codes, or who pays for a tutor, or who simply sits next to the kid, will keep building things at home. The families without that will meet AI a year later as a machine that answers questions. That is not an access gap. It is a tool-versus-answer gap, and it does not close on its own.
Mandate. Make AI a subject from age five, train every teacher, publish the curriculum. This is more serious than a ban, and the UAE deserves credit for treating it as a national priority rather than an IT ticket. But a mandate is a statement about access and instruction, not about design. A weekly AI class where a child watches a demonstration and answers questions about it is exactly the "performance without learning gains" pattern the OECD describes. Mandating access does not guarantee the thing that matters any more than banning it does.
Design. Put the child in the director's seat. The AI makes nothing the child did not ask for; the child looks at what came back, decides whether it is right, and sends it back when it is not. The output has to be a thing the child made, a game, a story, a quiz, a tool, so there is no answer to copy. The session is bounded. An adult can see every word of it. This is not a third policy option so much as the missing question inside the other two: when a kid and an AI are in the same room, who is doing the thinking?
What design looks like from the kid's side
We can describe this concretely because we watch it every day, at small scale, in our own product. When a nine-year-old asks for "a game where I'm a fox who delivers mail in a snowstorm" and a playable game appears a minute later, the interesting moment is not the appearance. It is the next thing the kid says.
In the first roughly 70 kid revisions we classified on Xyplor, 55 percent were what we call fixes: the child rejecting what the AI produced and redirecting it. Not adding, not starting over. "No, that's wrong, do it differently." We wrote up that number, and its limitations, in an earlier post. It is a small sample and a rough proxy. But it is a proxy for exactly the thing the OECD is worried about: whether the human in the loop is doing any cognitive work. A kid who rejects the machine's output more than half the time is not outsourcing. A kid copying an answer into a worksheet is.
That is the design principle in one sentence: the child should be able to be wrong about the AI, and to correct it, and to see the result of the correction. Ban and mandate are both silent on it.
What a parent can do this month
If your kid is in a district that just switched AI off, the ban has handed you the job. The useful version of it is not "let them use ChatGPT" and it is not "keep them away from it." It is: pick one bounded thing to make together. A story. A quiz about their favorite animal. A game with rules they invent. Let them direct the tool. When it gets something wrong, and it will, do not fix it for them. Ask them what is wrong and let them send it back. Twenty minutes, with you in the room or able to read every word afterward. That is the whole curriculum.
If your kid is in a system that just mandated AI, ask the school one question: in the AI class, what does my child make? If the answer is "they learn about AI," push for the version where they make something with it and have to judge the result. The OECD gives you the citation.
And whatever the district decides, the screen-time question is worth answering the same way on both sides. New York's 30-to-45-minute guidance is reasonable. A bounded session that ends with a made thing is a better unit than a minute count, but it is not an argument against limits. It is a better limit.
About us, in one paragraph
Xyplor is an educational website for children ages 6 to 17 and their parents. Kids make games, quizzes, tools and storybooks by directing an AI; every prompt and every result is visible to the parent, and nothing is published without a parent's approval. We are a small team, our data is early, and we are one option among several. Our position paper on student-centered AI policy is CC BY 4.0 and says all of this at length, with the argument for age bands and against blanket bans written in May, before this week made it timely.
What we do not know
We do not know how the New York or Los Angeles reviews will land, and it is possible both come back with something thoughtful. We do not know whether the UAE curriculum is design-heavy or demonstration-heavy in practice; the published outline includes "innovation and project design" as one of seven domains, which is promising, and a domain name is not a classroom. We do not have longitudinal data on our own kids, and we will not pretend to. What we do know is that the two biggest districts in America and one entire country answered the same question this week, and the OECD's evidence says they all answered a different question than the one that matters.
Sources: NYC Mayor's Office release and Chalkbeat coverage (Sept 1, 2026); LAist, K-12 Dive and the Washington Post on LAUSD (Sept 2-3, 2026); Gulf News on the UAE curriculum (May 2025) and the 2026-27 teaching plan (Sept 2026); OECD Digital Education Outlook 2026.