The headlines this week landed gently, given the political weight behind the policies they described. Schools across the country had pulled student phones into magnetic pouches and locked desk drawers, hoping a quieter classroom would translate into better test scores. A new working paper analyzing more than 4,600 schools from 2019 through 2026 found the effect on test scores was, in Duke professor E. Jason Baron's words, "basically close to zero" (Fortune, May 2026). Bullying didn't drop. Attendance didn't budge. The bans worked at the thing they were narrowly designed to do, which is to say students looked at phones less. The things schools hoped would follow, almost none of them moved.
This is the moment, weeks before districts finalize their summer policy revisions and back-to-school AI guidance, to take the lesson seriously. Phone bans were the closest thing this generation of school leaders had to a clean technological intervention: identify the device, remove it, watch the cognitive benefits accrue. That theory of change just took an empirical hit. AI policy is being written now on something close to the same theory, and AI is a harder problem than phones in almost every way that matters.
The bans worked at exactly the thing they were designed for
It helps to look at what did change. The share of teachers who reported students using phones in class for personal reasons fell from 61 percent to 13 percent, and GPS pings inside school buildings during the school day dropped by roughly 30 percent by the third year of pouch use (Fortune, May 2026). Teachers reported relief from the day-to-day burden of phone management. Jonathan Haidt, a vocal proponent of bans, described phone-free schools producing "more social interaction in class, and a lot more noise and laughter in the hallways, and at lunch" (Fortune, May 2026). The Scientific American write-up of the same analysis describes initial suspensions and lowered well-being scores in year one giving way, by year three, to higher self-reported well-being and teacher satisfaction (Scientific American, May 2026). Bans calmed classrooms. They didn't move tests.
That gap, calmer classrooms but no test-score lift, is the part most districts haven't fully sat with. The phone was a real distractor. Removing it removed real noise. But test scores are not a measurement of noise. They are a measurement of what gets learned inside the available time. Removing a distraction creates time. It does not, on its own, fill that time with the kind of thinking test scores reflect.
You can subtract a distraction. You can't subtract your way into learning.
There is a version of the phone-ban story where everything works as predicted: students aren't on phones, so they pay attention to the lesson, so they learn the lesson, so test scores go up. The mechanism is intuitive enough that it underwrote a lot of policy. What the new data suggests is that the second-to-last step was the load-bearing one, and we built around it. Students freed from phone use don't automatically attend to instruction. They have to be given something worth attending to, taught how to attend to it, and held to the cognitive work the attention enables. Removing the phone is necessary, but nowhere near sufficient.
"Average effects on test scores are basically close to zero." (E. Jason Baron, Duke University, in Fortune, May 2026) A separate NBER analysis of one large Florida district by Figlio and Özek did find modest positive effects, on the order of 1.1 percentiles overall and 1.4 percentiles for male students, with about half of the gain explained by improved attendance (Figlio & Özek, 2026). That is a real result, and it is the result schools should hope to replicate. It is also the ceiling of what restriction alone seems to buy in the best-implemented cases, on a metric that conventional school improvement plans regularly move further with less expense. The 2013 UK study by Beland and Murphy at the London School of Economics, often cited in favor of bans, found a 6.4 percent test-score gain after phone bans (Beland & Murphy, 2013). It landed in a different world. Phones in 2013 were dumber, social media less consuming, and the assignments students were being asked to complete looked roughly the way they had for decades. The 2026 data is the system's honest answer about what the same intervention does today.
AI is not a phone
The standard move from here is to argue, by analogy, that schools should "ban" AI the way they banned phones. The analogy collapses on inspection. Phones are a parallel distractor. They pull attention away from the work, but they don't do the work. AI is substitutive. A student with a phone in their pocket is failing to write their essay. A student with ChatGPT open in another tab has an essay being written.
This distinction is not theoretical. The Center for Democracy and Technology's nationally representative survey from October 2025 reported that 85 percent of teachers and 86 percent of students used AI during the 2024-25 school year, with half of students reporting school-related use (CDT, 2025). Seven in ten teachers said they worry AI weakens skills students need to learn. Only one in ten teachers reported any training on how to respond when a student's AI use seems harmful. The base rates make a clean ban a fantasy. The training rates make any other strategy harder to execute.
Consider what it would take to ban AI the way phones were banned. Phones are objects; AI is a capability that lives inside every device the student already uses. Phones leak around bans because students hide them. AI leaks around bans by existing inside the laptop the school provided. The pouch model has no analog. If a school nonetheless succeeds at restricting AI use during the school day, the assignments most affected, which are the multi-day essays, the research projects, the problem sets done at home, are exactly the assignments AI is still available for, off-site, after the bell.
What the phone-ban evidence actually argues for
Read carefully, the new study does not argue for keeping phones in classrooms. Well-being improved. Teachers got a meaningful quality-of-life gain. The classrooms are noticeably calmer. Those outcomes matter on their own terms. The study's argument is narrower: if your case for the ban was that test scores would rise, you bought a different intervention than the one you ran. The behavioral and well-being gains were the real return.
For AI, the same logic suggests something specific. Restricting AI access during the cognitive moments where thinking is supposed to happen, drafting, reasoning through a problem, struggling with a hard reading, is reasonable on the same grounds. It clears space. It does not, by itself, do the teaching. If the only intervention is the restriction, the test scores won't move, and they might move down once students have spent years offloading the thinking the restriction is now asking them to do.
A growing body of research suggests the offloading is already underway. Gerlich's 2025 study in Societies, drawing on a survey of 666 participants across age groups, found a significant negative correlation between frequent AI tool use and critical thinking scores, with cognitive offloading mediating the relationship; younger participants showed both higher AI dependence and lower critical thinking scores (Gerlich, 2025). A separate quasi-experimental study reported the opposite when AI was used to delegate lower-order tasks under explicit instructional scaffolding, leaving students to do the analysis and evaluation themselves. The split between those two findings is the entire point. The tool isn't the variable. The instruction wrapped around the tool is the variable.
The instruction is the policy
The temptation, for any administrator staring at the phone-ban data, is to read it as proof that interventions don't work and quietly walk back. That would be the wrong read. The interventions that did work, at the things they were designed for, show up clearly in the same data. The places where the analysis disappoints are the places the original theory of change overpromised.
For AI in schools, this argues for a few moves that look obvious in retrospect and almost no district has actually made. First, assume the base rate. Roughly nine in ten students and teachers are already using AI. Policies that begin from a premise of non-use are policies that will be ignored. Second, write assignments that require visible thinking. Process-visible writing, meaning drafts saved with timestamps, comments and revisions logged, oral defense of choices the writer made, is the AI-era equivalent of asking a math student to show their work. It is the part of the assignment AI can imitate but not replace, because the thinking that produced the writing is what is being assessed, not the writing alone. Third, train teachers, and train them on the specific question of when AI use helps and when it doesn't. The CDT survey's one-in-ten training rate is the policy lever everyone is ignoring.
What this looks like on a Monday morning
A working teacher cannot redesign every assignment between now and August. The useful narrowing is to identify one or two assignments per unit where the cognitive work is non-negotiable. A synthesis essay. An argumentative response to a hard text. A multi-step problem set with a written explanation. Rebuild those assignments around visible process. Keep the other assignments. Let AI assist where assistance is fine. Make the high-stakes pieces about the thinking, and make the thinking observable.
An administrator can do something narrower still. Audit the AI policy currently on the books and ask what theory of change underwrites it. If the theory is restriction-as-cognitive-intervention, the phone-ban evidence is now arguing against it. If the theory is restriction-as-classroom-management, the evidence supports it but does not promise test-score returns. If the theory is instructional, the policy needs an instructional companion document, and that document does not yet exist in most districts.
A parent can ask the school a single useful question. What does my child's school require them to do, this year, that a chatbot cannot do for them? If the answer is nothing, the policy is the assignment, and the assignment is what needs to change.
A final thought
Phone bans were the most popular school technology intervention of the last decade. They had broad bipartisan support, enthusiastic teacher buy-in, and most of the major news outlets nodding along. The newest, largest study of their effect says the thing schools hoped would happen mostly didn't happen. That is not an indictment of the bans. It is an indictment of the theory of change underneath them, that removing a thing and waiting was, by itself, an instructional plan. Schools are now being asked to take the same posture toward AI, by the same voices, with the same theory. The case for caution is not nostalgia for the bad old days of phones. It is the data.
Sources
- Fortune. "'Close to zero': Schools are spending tens of millions banning phones from classrooms, but test scores aren't improving." May 8, 2026.
- Scientific American. "School cell phone bans may boost student well-being but not test scores, new study suggests." May 4, 2026.
- Washington Post. "School cellphone bans don't affect test scores or attendance, study finds." May 4, 2026.
- Figlio, D. N. and Özek, U. (2026). "School Cell Phone Bans and Student Achievement." NBER Working Paper 34388.
- Center for Democracy and Technology. "Hand in Hand: Schools' Embrace of AI Connected to Increased Risks to Students." October 2025.
- Gerlich, M. (2025). "AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking." Societies, 15(1), 6.
- Beland, L.-P. and Murphy, R. (2013). "Ill Communication: Technology, Distraction & Student Performance." Centre for Economic Performance, LSE Discussion Paper 1350.