Should schools ban AI?

A blanket ban is not what the largest school system to act has done, and the evidence does not yet support either a ban or a rollout. New York City announced on 2 September 2026 a one-year moratorium on student-facing generative AI from 2-K to 8th grade, affecting nearly 600,000 pupils, while keeping teacher use, supervised high-school pilots and exceptions for assistive technology. NPR reported on 28 September that a Stanford review of more than 800 papers found the research on how AI helps students extremely limited. RAND found only about a third of students saying their school has rules at all. The decision a school controls is narrower and more useful than ban or allow: which work pupils do unaided, who decides when a tool comes in, and what happens when a teacher suspects misuse.

What New York decided, in its own words#

The Mayor’s Office release of 2 September sets out “a one-year moratorium, effective in the 2026-2027 school year, on student-facing generative AI for children in 2-K through 8th grade”, a decision it says “will impact nearly 600,000 public school students, or two-thirds of the system’s total enrollment.” It is less than a ban in four ways. Teachers “will continue to be permitted to use AI for instructional planning and operational tasks”. Exceptions cover assistive technology for students with disabilities, multilingual learners and career programmes. Up to 50,000 high-school students join five pilots, each “under the direct supervision of a trained educator”. And a coalition of students, teachers, parents and others will assess the year and publish recommendations. One part is absolute: “Companion chatbots will be prohibited across all grades.” The Mayor, Zohran Mamdani, gave the reason as a claim about how children learn: they need to “build relationships with educators and wrestle with tough problems on their own”. Lee V. Gaines at NPR, read here at WWNO, reported that Los Angeles Unified has since barred students from using AI on district devices.

The case against a moratorium#

Erin Mote, chief executive of the non-profit InnovateEdu, put the opposing case to Alyson Klein at Education Week on 25 September, and she speaks as someone who once banned social media in her own school: “I do not think AI moratoriums are helpful. I think that’s the mistake we made with social media.” Her argument is that pupils will meet the tools anyway and that school is the safer place to meet them. She is not relaxed about the products. “It’s abundantly clear that the tech companies can’t self-regulate”, she said, and on one category she agrees with New York: “You should ban AI parasocial companions.” Her working rule for pupils is the one the estate calls human at the start: “you should never do your first draft in AI.” Robin Lake of the Center on Reinventing Public Education told NPR the present position “feels very Wild West in terms of the randomness” and that the questions are bigger than a school or district can settle alone.

What the evidence does and does not say#

NPR’s summary of the Stanford review is that research on how schools can use AI to help students and avoid harm “is extremely limited”, with some studies suggesting “that carefully designed AI tools show more promise than general-purpose chatbots.” Rebecca Winthrop of the Brookings Institution drew the conclusion for practice: “I don’t think you should just sort of wildly experiment, especially with really marginalized communities”. That matches what the estate has graded. A tutor built to withhold the answer can teach, as the page on AI tutors records, and an open chatbot used for answers raises marks while the tool is there and lowers them when it is gone, the finding at the centre of whether children should use AI. Neither result tells a head whether to ban. Both say the unit of decision is the task and the design, which a ban and a rollout equally ignore. What governments have published is on official guidance on AI in education.

What pupils say is happening#

Melissa Kay Diliberti of RAND, in an interview published on 30 September, described students who use the tools and distrust them at once. In RAND’s youth panel of December 2025, 67 per cent endorsed the statement that the more students use AI for schoolwork, the more it will harm their critical thinking. On rules: “Only about a third said there were schoolwide rules. A lot of kids said they didn’t know.” A second report shows what fills the gap. The Center for Digital Thriving at Harvard, in An AI Policy Isn’t a Playbook, found that when teachers and principals described an AI dilemma in their own words, 74 per cent of the teachers’ accounts and 69 per cent of the principals’ were about cheating. Its interviews with 31 teenagers found something adults had not been looking for: students who “cap their effort, or perform mediocrity on purpose to not get flagged”. One told the researchers: “We can’t make ourselves sound too smart anymore”. Beck Tench, one of the authors, gave it a name in Education Week on 30 September, “cognitive withholding”, done “to evade being falsely accused in environments where accusations don’t get any kind of conversation.” The estate’s evidence on why accusation by software fails is on whether AI detection works.

Who decided, when the tool arrived anyway#

A ban assumes the school chooses what comes through the door. Grace Jiang reported at Education Week on 1 September that Google Classroom had “recently opened use of its Gemini chatbot to K-12 students of all ages, giving schools little time to determine how to restrict or monitor its use.” The detail is fairer to Google than the headline: the feature appears only where a district had already given students access to Gemini, it stays off where it had been switched off, the company announced the plan in June, and administrators can turn it off at any time. Robbie Torney of Common Sense Media still described the approach as one that “defaults to adoption rather than permission.” A principal quoted in the piece, Shameka Joyner, named what she wanted protected: “I don’t want teachers to lose their creativity, their professional judgment”. A setting chosen by a supplier months earlier is a decision about a school’s pupils, and somebody at the school should know they made it.

The decision a school controls#

The pattern in Rules Before Tools fits a school as well as a firm. Which work a machine may do: written by task and by age, with the work that forms a skill done unaided first, which is New York’s reasoning without the blanket. Who can bring a tool in and who can switch it off: the head and governors, by name, and never a default. What people must remain able to do: pupils, the reading, writing and reasoning the tool can imitate; teachers, the marking and the judgement of a pupil’s work, examined on how AI will change teaching. And how anyone would know it had gone wrong: a conversation with the pupil before any accusation, and assessment that shows the process. James Wells of Middlesbrough College wrote at FE Week on 1 October that the best way to stop AI replacing learning “is through assessment redesign.” A school that has written those four answers has a policy that survives the next product. A ban on its own is one answer to the first question.

What this does not show#

No study compares schools that paused AI with schools that did not, and New York’s year has only begun. The Los Angeles decision is one clause in NPR’s report. The Stanford review was read through NPR’s summary and not at source. RAND’s figures are self-reports from 1,214 young people aged 12 to 29 in the United States. The Harvard percentages are shares of the dilemmas educators chose to describe, 171 from teachers and 85 from principals, and the finding about withheld effort rests on 31 interviews. “Cognitive withholding” is a phrase one author used in an interview; it does not appear in the report. Nothing on this page is UK evidence, and it says nothing about the Department for Education’s position, which sits on the guidance page.