New York bans generative AI for 600,000 schoolchildren while universities opt for nuance, not vetoes

By Zendoric · September 5, 2026.

New York has just imposed a one-year moratorium on the use of generative AI by students from pre-K through eighth grade, a measure that affects some 600,000 pupils and is combined with new limits on screen time, the city's school district has announced. A small group of upper high school students will have controlled access, with mandatory AI literacy classes and supervised pilots.

Los Angeles Unified, the country's second-largest school district, has moved in the same direction: it bars its students from using generative AI on the devices the institution itself provides during the 2026-27 school year, while it designs a broader policy; teachers will retain access. Florida, for its part, is debating a rule that would require all of its public institutions —from K-12 to university— to adopt explicit AI use policies, according to The EDU Ledger.

The contrast with higher education is telling. A 2025 study published in the International Journal of Educational Technology in Higher Education, which analyzed the top 50 U.S. universities, found that the vast majority avoid a single campus-wide rule: 94% of the institutions studied had faculty guidance insisting on setting and communicating clear expectations, but left it to each professor to decide whether AI is permitted, restricted or banned on a given assignment. Columbia Law School, for example, imposes no blanket ban in its 2026-27 policy —it allows AI to be used as support or to stress-test ideas, but prohibits it from replacing the student's legal reasoning or being used in exams— while UC Berkeley's School of Law flatly prohibits using AI to conceptualize, outline, draft, edit or translate any graded work.

Behind these policies lies a technical problem that is rarely explained clearly enough: AI detectors do not work reliably. A June study comparing four widely used detection tools found substantial differences in their ability to identify AI-generated or hybrid texts, and concluded that they should not be used as the sole evidence in high-stakes academic decisions. A paper published in August in ScienceDirect, which examined 13 detection systems on student papers, theses and code, found systematic failures and concluded that their performance remains inadequate for high-stakes assessment. And a 2023 Stanford study —already well known but still the reference point on the problem— found that seven widely used detectors wrongly classified as AI-generated an average of 61.3% of essays written by non-native English-speaking students.

That last figure matters more than it seems: it means that relying on automated detection to punish 'AI cheating' can disproportionately penalize those who already start at a linguistic disadvantage, rather than those who actually cheated. It is exactly the kind of short-term friction we have been documenting for months in other areas —algorithmic biases do not distribute error neutrally— and here it lands squarely on the group least able to defend itself before a disciplinary board.

Our reading is that what we are seeing is not a rejection of AI, but the hangover from an adoption carried out too fast and without guardrails. Over the past few years, schools and universities rushed to bring generative AI into teaching and administrative tasks before it was clear what could and could not be done with it. Now that the technology has been in classrooms long enough for its effects to show —and New York's own announcement acknowledges that concern about learning was a central reason for the measure— the system is correcting. It is the same pattern we already saw in the corporate world with the premature AI-driven cuts that companies such as Klarna or Commonwealth Bank had to reverse: adopting first and understanding the consequences later is expensive, and backtracking is more common than the triumphalist narrative about AI suggests.

The underlying distinction that runs through all these policies —and the soundest one in the debate, in our view— is not whether a student used an AI, but whether their work demonstrates what the institution intended them to learn. A nursing student needs their own clinical reasoning before an AI helps them organize information; a future lawyer can use AI to stress-test an argument, but must be able to build and defend it themselves. That logic fits what we already argue about employment: in a market where AI will be ubiquitous, the winner is neither the one who avoids it nor the one who delegates to it without judgment, but the one who learns to orchestrate it without losing the judgment that makes it useful. The schools now slamming on the brakes are not betting against that future; they are buying time so that their students arrive at it knowing how to think, not just how to ask.

The risk, in the medium term, is that the gap between K-12 caution and university pragmatism translates into generations who reach university without the basic AI literacy the labor market already demands, or that Florida and other states turn a reasonable governance need into a rigid mandate that draws no distinction between ages or subjects. It is worth watching closely whether these moratoriums become permanent frameworks for AI-assisted learning, along the lines New York's pilots suggest, or simply freeze in place through bureaucratic inertia for years.