Mayor Zohran Mamdani and Chancellor Kamar Samuels announced that New York City public school students may not use student-facing generative AI for the 2026-2027 year. The rule reaches from 2-K through 8th grade and covers nearly 600,000 children, about two-thirds of the system. High schools receive a tightly rationed pilot, capped at 50,000 students citywide and no more than five classes per building. In a system of roughly 280,000 high schoolers, most of them get nothing. Every high school student will instead sit through two 45-minute modules on AI doomer and anti-data center propaganda each year. Companion chatbots are barred in every grade. Teachers may still use the same tools to plan lessons and grade papers. The mayor and the chancellor announced all of this under the banner of putting students first.
Picture a child who adds fractions by adding the denominators. A competent tutor notices the error, shows why the shortcut fails, offers a picture, asks her to try again, and steps back once she has it. If her brother finishes early, the same tutor hands him a harder problem instead of telling him to sit quietly. That is ordinary differentiated instruction, the thing a classroom of 29 cannot reliably do and a family of ordinary means cannot buy by the hour. The city has now told that child she may not have the machine version of that help at school for a year.
Mamdani’s justification at the press conference was blunt. “I have yet to see a study showing that AI is beneficial for students in elementary and middle school.” Mandani may not have looked, but the research is plentiful. Stanford’s education researchers maintain a repository of work on AI in K-12 schooling, and by March of this year it held more than 1,100 studies. From that pile, the Stanford team pulled 20 rigorous causal studies for close review. The pattern is the ordinary one in educational research: real gains under careful design, weaker results under sloppy design, and a reminder that the tool matters less than how a school uses it. It does not argue for a citywide prohibition.
The newer work is harder to wave away on grounds of age or sponsorship. In August, economists at the University of Toronto released results from a two-year randomized experiment in 18 Tennessee middle schools, the exact grades New York just closed. Students given Khan Academy with its Khanmigo AI tutor gained about 1.3 national percentile ranks per term. A companion experiment with more than 6,000 middle schoolers found that adding AI assistance to a math practice platform improved students’ ability to recover from their own mistakes. The funders were the Smith Richardson Foundation, Canadian public research councils, and J-PAL at MIT. No chatbot company wrote the check.
The evidence is not limited to adolescents. A randomized study of more than 1,000 students in grades 3 through 6 gave their human tutors an AI copilot that suggested better explanations in real time. Mastery rose 4 points overall and 9 points for children stuck with the weakest tutors, at a computing cost of roughly $20 per tutor per year. In North Carolina, a randomized trial across 63 middle schools and nearly 6,000 students found that structured computerized math feedback raised state test scores by a tenth of a standard deviation a full year after the program ended. In Nigeria, a six-week after-school program using a general-purpose AI assistant produced a 0.31 standard deviation gain, measured on a test the students took without the AI in front of them. At Harvard, a purpose-built AI tutor more than doubled the learning gains of the university’s celebrated active-learning physics classes. Some of these are working papers. Some involve older students. All of them exist. The mayor said he had not seen a study that showed benefit, but I wonder if he bothered to look. He had seen a reason to stall.
The reason the stall costs so much is older than any chatbot. In 1984, the psychologist Benjamin Bloom showed that a child with a personal tutor performed better than 98% of children taught in a conventional classroom. He called it the two-sigma problem because nobody could afford a tutor for every child. A conservative does not need to treat a language model as magic to notice what has changed. A tool that can teach at the pace of the learner, at odd hours, at near-zero marginal cost, is the first plausible mass answer to Bloom’s constraint. It still has to be tested with adults in the room and with exams the student takes unaided. New York made that test illegal for most of its children.
The politics of the decision are not mysterious. Mamdani is a member of the Democratic Socialists of America. The most powerful institution in New York education is the United Federation of Teachers, and the UFT’s parent union, the American Federation of Teachers, has already told the country what it wants from the technology. In July 2025, AFT president Randi Weingarten announced a $23 million academy, built with Microsoft, OpenAI, and Anthropic, and housed at UFT headquarters in Manhattan, to train 400,000 teachers, declaring that, with it, “teaching and learning can be enhanced.” Teachers may use the tools to draft a unit and mark a stack of essays. The child sitting in front of the teacher may not use them to learn, which is a rule that protects the teacher’s workflow and denies the student a chance to catch up.
Los Angeles made the same move in the same week. Los Angeles Unified, the second-largest district in the country, blocked generative AI on district devices for students in every grade while leaving teacher access in place. Other large districts are reviewing their rules on the same timetable. A pause can be dressed as prudence. In a school year, it is a 7th grader’s lost instruction, and a 7th grader does not get the year back. Families who can pay will route around the rule. Alpha School’s New York campus, which builds its day around AI-driven personalized instruction at 180 Maiden Lane, charges $65,000 a year. The professional-class child will have the tutor at home whether or not the chancellor permits it at school. The child in a Bronx public school will not. A movement that campaigns in the name of the working class has, in practice, arranged for the working class to learn more slowly than the rich.
The replacement curriculum is the part of the policy that will linger after the year is up. Those two mandatory 45-minute modules are not instruction in how to use a powerful tool, how to check its answers, or how to build something with it. By the city’s own framing, they are lessons in caution and harm. New York banned companion chatbots in every grade, then wrote an hour-and-a-half of high school time whose premise is that the machine and the data center are to be feared. Beijing mandated at least 8 hours of AI education per year for every primary and secondary student beginning last fall, with the work folded into assessments. President Trump’s April 2025 executive order made AI literacy and teacher preparation a national priority on the theory that early exposure “sparks curiosity and creativity.” One government is teaching children to build. New York is teaching them to flinch, and it is doing so on a separate track from the teachers who will receive the $23 million academy downtown.
The standard reply is that the science is young and that caution costs nothing, but every part of that reply is wrong the science is not young it is 1,100 studies deep and growing by the month and caution is the most expensive habit in education because its cost is paid in years a child cannot recover and the people demanding more study will demand still more study when the next round of results arrives because delay was the product. Nobody serious proposed putting an unsupervised chatbot in a kindergarten. What the research supports is supervised, evaluated use, with tests taken unaided so the learning can be seen. That is the experiment Mamdani closed to 600,000 children.
The conservative program is narrower than the rhetoric around it. Let AI into the classroom under a teacher’s eye. Measure what children can do afterward without it. Publish the results, cancel what fails, expand what works, and tell parents the truth about both. Give the public-school child the same kind of paced help the private-school child already buys. AEI’s John Bailey told the Senate last fall that AI “places expertise within reach of anyone willing to learn and apply it.” Expertise for everyone has been the promise of public education since Horace Mann. It has not been this close to a practical form.
Lincoln told Congress in 1862 that as our case is new, we must think anew and act anew. The American child’s case is new in one respect only. For the first time a patient expert can sit at her elbow in every subject, at hours a building cannot staff, at a price a district can bear. New York considered that fact and reserved the tool for union members. Parents in the Bronx cannot buy the Alpha product on Maiden Lane. They can demand that the public school stop treating a $20 tutor-copilot as contraband, measure the child without the machine in front of her, keep what raises the score, and kill what does not. Mamdani chose the pause. That choice is now the record.
If you enjoy my work, please subscribe https://x.com/amuse/creator-subscriptions/subscribe
Alexander Muse is a Fellow at the John Milton Freedom Foundation and publishes daily political analysis at amuseonx.com. Primary sources cited in this piece are linked inline; campaign finance figures are drawn from FEC filings, polling data from publicly released crosstabs, and legal claims from filed pleadings. Corrections are posted to the original URL with a dated changelog. Readers who identify errors are invited to contact the author directly. Each op-ed edited for grammar and clarity using Ai in partnership with Grammarly. Data provided in a sponsored partnership with Polymarket.




And there are no studies showing islamo communists destroy cultures either. Just abundant physical evidence
Communism MUST keep citizens as dumb as possible.
Notice I did not say uneducated.
The most educated in America, today, sure seem to be the dumbest.