On September 2, as students returned to class, New York City implemented a one-year pause on generative AI use for nearly 600,000 students across pre-K, elementary, and middle school. Los Angeles soon followed, blocking it for every student on a district-issued device. Together, the country’s two largest school systems came to the same conclusion: Better to wait a year than to keep putting untested tools in front of kids.
Neither district is calling the decision permanent. Both are responding to parents who worry that untested AI tools are reaching students before anyone knows whether they improve learning. New York City Schools Chancellor Kamar Samuels said it best: “We’re standing firmly in our belief that innovation does not mean more technology, and over the next year, we will lead with evidence to make sure technology serves learning—not the other way around.” 
That leaves an obvious question: Where will that evidence come from?
ADOPTION IS OUTPACING EVIDENCE
Teachers, school leaders, parents, and students are making consequential decisions about AI without reliable information about what works, for whom, and under what conditions.
Sixty percent of teachers use AI for work, yet 82% report they’ve received no formal guidance on how to use it. Just half of middle and high schools have any AI policy, and only 6% of teachers in those schools say those policies are clear. This lack of clarity shows up in how people feel about AI, too. Thirty-one percent of Gen Z now say AI makes them angry, up from 22% a year earlier, and 80% believe relying on it will make learning harder.
Still, banning AI in school won’t take it out of students’ lives completely. A recent Common Sense Media survey found that 91% of teens ages 13 to 15 use AI for schoolwork or homework. And with 21% of adults already using AI at work, students need to consider the growing role of AI in their future careers.
It makes sense that school district leaders, educators, and families want to know when AI adds to student learning and when it takes away from it. But even with the best intentions, it’s hard to write good policy, train teachers, or make smart purchasing decisions when the evidence underneath is this thin.
WHAT WASHINGTON SHOULD DO
Congress and the federal administration have a chance to put the existing infrastructure to work and provide the evidence we need to safely and confidently advance student learning. The Alliance for Learning Innovation worked with AI researchers, practitioners, and our coalition of more than 150 members to develop federal recommendations on this exact topic. Four main takeaways:
Pass the NEED Act. This past September, Senators Michael Bennet (D-CO), Thom Tillis (R-NC), and Chris Van Hollen (D-MD) reintroduced the bipartisan NEED Act, which would create a national center at the Institute of Education Sciences (IES) to develop and test bold new approaches to teaching and learning.
The model comes from the Pentagon’s Defense Advanced Research Projects Agency (DARPA), which helped build the early internet and laid the groundwork for Siri. Like DARPA, the new center would be run by expert program managers who can quickly move money to promising and high-impact ideas, invest more when they show results, and pivot when they don’t.
Study AI tools with robust data. One example of how to study the issue: New York City plans to run AI pilots in a small number of high school classrooms this year to learn what’s working for its own students. The federal government should help school districts nationwide do the same.
Where pilots aren’t feasible, a coordinated national network of testbeds, combining real classrooms with synthetic environments that simulate how students use AI, would let researchers test tools across diverse schools and communities. The result would be practical evidence about how AI holds up in the everyday reality of teaching, plus clearer signals for developers and investors about which tools deserve to scale. Federal investment in these testbeds would give schools a shared, safe way to judge what is and isn’t working in classrooms.
Measure what students learn. Many current evaluations can’t tell the difference between a student’s output and AI. Federally funded AI studies should measure how students perform with a tool, then evaluate them again on the same topics later without it. The American Psychological Association made the same case, urging that any AI-powered learning tool be judged by what students can do after they stop using it.
This is exactly the kind of rigor the Department of Education’s own guidance and a growing number of outcomes-based contracting models (which tie vendor payment to results), are asking for. Currently, few evaluations consistently meet this standard.
Give school districts a common evaluation tool. Districts are evaluating AI tools without a clear definition of what makes it safe, effective, and ethical. The National Institute of Standards and Technology (NIST) already published a voluntary framework for AI Risk Management that organizations use across industries. IES should partner with NIST to build an education-specific version, aligned with the EDSAFE AI Alliance’s SAFE Benchmarks, that gives states and districts shared standards for judging tools on learning impact, safety, privacy, and transparency before those tools reach classrooms.
THE CLOCK IS TICKING
New York City was clear that this moratorium lasts one year for a reason. The city said it will spend that time studying how AI affects students, and Los Angeles and other districts have set similar deadlines.
That study shouldn’t happen one district at a time, and it shouldn’t be limited to the contexts of large and urban school districts. If we invest in our federal research infrastructure now, we can build a trustworthy evidence base across the country so that all students walking into class in Fall 2027 have access to tools that we know work.
Sara Schapiro is executive director of the Alliance for Learning Innovation.Â
