School AI Rules Splinter as NYC and LA Ban Most Student AI Use

A school year that opened with bans

The 2026-27 school year started with a clear signal from the two largest school districts in the United States: for most students, generative AI is off-limits. New York City Mayor Zohran Mamdani announced on Sept. 2 a "moratorium" on student-facing generative AI. Students below high school are covered by the moratorium, and companion chatbots are banned in every grade.11 Under the city's plan, about 5% of high schoolers will join five AI pilot programs. All high schoolers will take AI literacy classes, and the city says it will use the year to study how these tools affect students.11 Mamdani pushed back directly on the industry's argument, saying the city does not accept that AI-powered early education is either inevitable or necessary.6 One day later, Los Angeles Unified said it would ban most student AI use on district devices this year.6

The big districts are not setting a national standard, though. Boston gives students supervised access to approved AI tools beginning in third grade, and its guidelines require students and staff to disclose any AI use.6 Denver has banned ChatGPT but lets students use MagicSchool, an AI platform built for education.6 What has formed is a patchwork that varies from state to state and sometimes from district to district. Much of the state guidance only recommends practices to districts rather than requiring them.11

Why the rules feel bewildering

The confusion is real, and it is not limited to younger students. A report published in late September found that fewer than 20% of teachers work at schools with a specific generative-AI policy.16 Families noticed the problem early. A year ago, educators were already telling reporters that no rule can keep students off ChatGPT on their personal phones and laptops.20 The stakes of having no rule are easy to see. One school leader put it this way: punishing a child for using ChatGPT when no policy existed is a problem in itself.20

In the absence of clear rules, students are drawing their own lines. A Westchester County fourth-grader offered a test that many adult policies are now trying to formalize: copying AI output is cheating, but using it to build on what you already know, in your own words, is not.11 Older students say the boundary is harder to locate. They ask whether summarizing readings or outlining with a chatbot counts as cheating, and some avoid asking teachers because admitting any AI use could mark them as suspects.15

In higher education, the trend is to replace campus-wide prohibitions with course-by-course rules. Templates published by Carnegie Mellon range from forbidding AI at every stage, including brainstorming, to encouraging it as long as students submit their prompts and the AI's responses.7 Duke lays out a similar range, from outright prohibition to free use with no acknowledgment required.8 The University of Chicago Law School applies a more principled default. If a task would be acceptable for another person or technology to perform, AI may perform it. If it would already count as plagiarism, it still does.5 Policy guides describe the main change in 2026 as a move from "AI is prohibited" to "undisclosed AI use is a violation."1 Most major universities now require disclosure and leave the details to individual instructors.2

The flexibility makes sense, but it has a cost. A single student can face several different rulebooks in one day, and that fragmentation is exactly what students describe as bewildering.

The research case for clarity over severity

An October study suggests the confusion is more than an inconvenience. In a survey of 863 Vietnamese undergraduates, students who saw AI rules as unclear, AI cheating as hard to detect, and institutional guidance as thin reported more AI-assisted cheating. That cluster had the strongest link in the model.14 Academic pressure, often assumed to drive cheating, showed no meaningful link on its own.14 The study has clear limits. Together, the factors explained only about 5% of the variation, and the authors say they cannot show that clearer rules would reduce cheating.14 Still, the direction fits what educators describe. Earlier research from three high schools found that self-reported cheating stayed fairly stable after ChatGPT arrived. Most students opposed having a chatbot write an entire paper but supported using AI to get started or to explain concepts.14

Taken together, these findings point somewhere specific. The case for clear, communicated boundaries is much stronger than the case for strict ones. A ban that students understand may work better than permissive rules nobody can interpret. A ban that pushes AI use underground may be worse than either.

That underground risk is a recurring concern. One researcher quoted in the September report warned that policies built on distrust create "secret cyborgs" who hide their AI use. She argued this can make struggling students "invisible" because they turn to a chatbot instead of a teacher.16 Policy guides for schools make a similar point, warning that blanket bans push use out of sight and make assessment disputes harder to settle.4

Detection, surveillance and false accusations

Enforcement is where integrity policies have done the most damage. Los Angeles Times reporting from California campuses described a UCLA sociology class where students were told to set up a mirror reflecting their entire desk during an online test. In another course, students had to keep their arms crossed or behind their heads during video exams.13 Students also complain about proctoring software that tracks keystrokes, eye movements or faces. Some now keep records of every draft in case they are accused.13

Many of those accusations rest on weak evidence. Last year AP reported on an English learner who translated his own work with DeepL, unaware that the tool had changed his wording, and was then flagged by an AI detector.15 Khan Academy founder Sal Khan has called false accusations a growing problem. He described a relative who deliberately added errors to their own writing so a teacher's detector would not flag it.20 Institutional guidance has largely accepted that detectors are unreliable. Faculty are now told to treat a detector score as one signal, compare it with a student's earlier work, and talk with the student privately before filing charges.2 Turnitin itself says its AI-writing model should not be the only basis for disciplinary action.1

The tutoring paradox

The academic-integrity debate sits beside a second question that may matter more: does AI help students learn? Here the evidence splits sharply, and the split explains much of the policy confusion.

Some results are strongly positive. In a randomized trial with Harvard physics students, a carefully designed AI tutor produced more than twice the learning gains of an in-class active-learning session, in less time, and students said they felt more engaged.22 Two randomized trials of live, chat-based math tutoring found that Google's LearnLM performed as well as human tutors at helping students fix mistakes. Students tutored with LearnLM later succeeded on harder topics 66% of the time, compared with 61% for students with human tutors alone.21 In a separate trial, a tool called Tutor CoPilot raised topic mastery by four percentage points, with the largest gains among students assigned to less experienced tutors.21 A macroeconomics trial found that a course-specific AI tutor improved exam performance, and the biggest gains came when the tutor was combined with group work.29

One of the largest studies points the other way. A randomized trial published this fall involved 2,379 undergraduates and 30 instructors at a large U.S. public university. Students given a course-integrated AI tutor earned final grades up to 0.37 standard deviations lower than peers in the same course without it.24 Participation on the course platform dropped even more, and the estimated grade losses for first-generation students were more than double those for continuing-generation students.24 The authors conclude that the tutor may have crowded out existing learning activities without improving performance.24

The best explanation for the contradiction is design and structure. An expert quoted by USA Today said research shows performance improves while students have AI and erodes once it is taken away. He also argued that answer-giving chatbots can remove the "productive struggle" that learning depends on, while more Socratic, narrowly focused tools keep students working through problems.11 The OECD's 2026 outlook describes the same pattern. General-purpose AI can improve the quality of student work while leaving little lasting learning, but AI built to tutor can raise outcomes. A 2025 meta-analysis also found larger gains when teachers guided how students used AI. The successful trials used purpose-built, structured tools, often with a human involved.2122 The large negative result shows that simply adding a tutor to a course does not guarantee those conditions.24

What this means for the rules

Read together, the reporting suggests the wrong question is driving policy. Asking "Is using ChatGPT cheating?" treats the tool as the problem. The research suggests the problem is AI use that replaces student thinking: either the student hands off the work, or the tool simply supplies answers. On that reading, New York's choice to pause and study, while teaching high schoolers AI literacy, is defensible given how mixed the evidence is. Denver's approach of blocking ChatGPT while allowing a purpose-built education tool follows the same logic.116

What no district or campus can defend is ambiguity. The policies most likely to hold up will say which uses are allowed, require disclosure when AI is used, and treat detector flags as reasons to ask questions rather than proof of cheating.13 Assessment will have to change too. AP reported that many educators already consider take-home essays an open invitation to cheat.15 Until assessments change and the rules become clear, students will keep guessing where the lines are, and some of them will be wrongly accused.