ChatGPT in Schools: 2026 Bans Clash With Mixed AI Tutoring Results

The pendulum swings back toward restriction

Generative AI in American classrooms looked very different at the start of the 2026–27 school year than it did a year earlier. The clearest sign came from the country's two largest school districts. New York City has placed a one-year moratorium on generative AI for public school students from 2K through eighth grade. The ban covers nearly 600,000 children, close to two-thirds of the district's enrollment.26 Mayor Zohran Mamdani said the district will also switch off or disable the AI features in more than 38 programs it had previously allowed, because they fall short of new safety and oversight standards.26 High schoolers are not fully exempt. Companion chatbots and mainstream services such as ChatGPT and Claude will be blocked at every grade level. Older students will get a limited set of approved tools and "AI critical thinking" courses.26 USA Today reported that only about 5% of high schoolers will take part in five AI pilot programs. All high school students will take AI literacy classes.23

Los Angeles Unified acted within a day. It said it would ban most student AI use on district devices this year while it reviews safeguards.2326 Mamdani also made a political argument: he rejected the tech industry's claim that AI-powered early education is inevitable or necessary.23

New York has been here before. Its education department banned ChatGPT on school devices and networks in 2023 and reversed the ban months later.26 This time the policy is framed as a study period. A new Technology in Schools Coalition of educators, parents and students will assess how the moratorium works and recommend what to do after it ends.26 The district also arrived at the ban by an uneven route. It released preliminary guidance in March 2026 built on a "traffic light" framework and promised a full playbook by June.1 One account found that, as of August, the district's own pages still described that playbook in the future tense.25

A patchwork, not a policy

Beyond the big cities, the picture is fragmented. Counts of how many states have official K-12 AI guidance vary by tracker. AI for Education counts 35 states plus Puerto Rico as of late August.1 Playlab's directory puts the figure at 40 states plus Puerto Rico as of September.4 The gap probably reflects different definitions of "official guidance" and the speed of new releases rather than a real disagreement. Either way, a large majority of states have now said something. In most cases that guidance consists of recommendations rather than mandates.23

The bigger change in 2026 is that some states have moved from advice to law. The PIE Network tracked nearly 100 state bills this year that could directly affect how students use AI.2 One legal review describes "no single state-law pattern":

• Ohio required districts to adopt an AI-use policy by July 1, 2026.3
• Oklahoma's SB 1734 took effect the same day. It requires educator review of AI output before that output is used for instruction or decisions affecting students. It also creates a parent opt-out with alternative instruction.3
• Maryland's AI Ready Schools Act requires every local system to name a central-office AI coordinator.6
• North Carolina's deadlines run into 2027 and 2028.3
• Florida issued policy in September that requires parental notification and opt-in or opt-out for direct student AI use, with a non-AI alternative for families who opt out.4
• Puerto Rico took the most prescriptive approach of all. It names Microsoft Copilot as the only authorized generative AI tool for students 13 and older.4

District choices vary just as much. Boston gives students approved AI tools under teacher supervision starting in third grade and requires disclosure of AI use.23 Denver bans ChatGPT but allows the education platform MagicSchool.23 Akron's board has considered banning ChatGPT and Claude in classrooms.28 Higher education has largely settled on a different model: students disclose their AI use and each instructor sets the rules for their course. K-12 remains more restrictive on average.21

Teachers are using AI without rules

While policymakers argue over student access, teachers have already adopted AI with little direction. A Gallup and Walton Family Foundation survey of 2,069 public school teachers found that about six in ten use AI for work. Only 18% have received any formal guidance on how to use it.9 Across ten tasks, 34% of teachers get no guidance at all and 48% get only informal guidance.9

The gaps are biggest where AI touches students directly. Sixty-nine percent of teachers say they have received no guidance on using AI for one-on-one instruction or tutoring. Fifty-eight percent say the same about grading.7 Even where guidance exists, it usually neither encourages nor discourages AI use.9 Schools in wealthier communities are somewhat more likely to provide direction than higher-need districts.7

The vendors are filling that gap. OpenAI expanded its free ChatGPT for Teachers program to 55 more school systems in 20 states. It says it now works with more than 100 K-12 organizations reaching over 300,000 educators, and the program stays free through June 2028.29 OpenAI also announced a data privacy agreement covering 16 states, which it called an industry first. The company says the program remains limited to educators.29 For older students, OpenAI began automatically placing under-18 users into a "ChatGPT for Teens" experience on August 18. That mode reportedly pushes back on shortcut requests and steers students toward Study Mode.25 These changes help explain why some districts are not using blanket bans. They are separating consumer ChatGPT from managed tiers like ChatGPT Edu that come with admin controls and data commitments.2227

What the research shows about AI tutoring

The policy fight depends on a simple question: does AI tutoring actually help students learn? The honest answer from 2026 research is that it depends on the tool, and the dependence is large.

The optimistic case relies on well-designed studies of purpose-built tutors. A 2025 randomized trial at Harvard, published in Scientific Reports, found that students using a custom physics tutor outperformed students in active-learning classrooms. The effect sizes ranged from 0.73 to 1.3 standard deviations, and the AI group's median time on task was 49 minutes versus 60 minutes in class.17 Stanford's National Student Support Accelerator summarized two newer randomized trials of AI built into live math tutoring:

• Google's LearnLM: supervising tutors approved 76.4% of its responses with little or no editing. Students it tutored succeeded on harder follow-up topics 66% of the time, compared with 61% for human-only tutoring and 56% for static hints.12
• Tutor CoPilot: this tool supports human tutors rather than replacing them. It raised topic mastery by four percentage points, and by up to nine points for students paired with less experienced tutors.12

The newest large trial points the other way. A randomized experiment at a large U.S. public university gave a course-integrated generative AI tutor to some sections and not others. It involved 2,379 undergraduates and 30 instructors. Students with tutor access had final grades 0.37 standard deviations lower in the exact-match sample. Their participation in course activities fell by 0.90 standard deviations.15 First-generation students lost more than twice as much as their continuing-generation peers.15 The grade drop appeared even though only about 15% of students offered the tool used it, averaging about four sessions. Students mostly asked it for direct answers rather than feedback on their own work.15

The authors place their results within a growing body of work. Several studies find that unrestricted general-purpose AI can raise scores on assisted homework while weakening later performance when students work on their own, a so-called "learning penalty." Guardrails built into dedicated tutoring tools appear to reduce that harm.15 They also caution that the Harvard result comes from one course at a highly selective university.15

Commercial claims rarely acknowledge this nuance. Vendor-adjacent coverage often repeats the Harvard numbers as evidence that AI tutors "doubled learning outcomes."14 Some of it cites figures such as "54% higher test scores" without clear sources.1718 More careful reviews make the narrower point. Gains tend to appear in structured subjects like algebra. Longer interventions beat short pilots. And AI tutors often tie with or slightly trail well-staffed small-group human tutoring.13

Reading the landscape

The research and the policy changes fit together better than the headlines suggest. The strongest positive evidence comes from tightly designed tutors, often with a human in the loop, used in structured subjects.1217 The warning signs come from open-ended, general-purpose access, where students can hand off the thinking itself.15 New York's policy follows that split. It blocks consumer chatbots, pauses AI for younger children, permits limited pilots, and allows teachers to use AI for planning but not for grading.26 Oklahoma's educator-review rule and Puerto Rico's single-tool mandate apply the same logic through law.34

The weak point is the middle ground. Most teachers still work without written rules, especially on tutoring and grading, which are the uses where the research is most contested.79 Until districts close that gap, the main question in classrooms will not be whether to allow ChatGPT. It will be whether anyone is checking how AI is being used with students.