In the past few weeks, New York City and Los Angeles have both moved to restrict generative AI in classrooms. Meanwhile, Miami-Dade has handed Google’s Gemini to more than 105,000 high schoolers. These are the most visible examples, but school systems of every size are confronting similar decisions even though research about which of these tools actually improve student outcomes barely exists.
In “The Evidence Base on AI in K–12: A 2026 Review,” researchers at Stanford’s SCALE Initiative worked through more than 800 academic papers on AI in K–12 education. Twenty produced strong causal evidence, meaning proof that the tool itself caused the change in learning. And none of those high-quality studies were conducted in K–12 settings in the United States.
That uncertainty is why Accelerate commissioned independent evaluations of AI-powered educational tools during the 2025–26 school year. Together, the studies followed roughly 21,000 students across fourteen states and tested the tools under real classroom conditions. Over the next few months, we will publish a synopsis of every study, whether the findings are encouraging, disappointing, or inconclusive.
Every district and provider featured in these reports agreed to be measured before knowing what the data would show. In a market that rewards confident claims over demonstrated results, that commitment itself is worth recognizing.
This month, EdReports provided a useful baseline for understanding how AI is moving through the instructional materials market, including the inconvenient observation that evidence for an established product rarely transfers to the AI features recently added to it. Instruction Partners studied 20 student-facing tools and talked with more than 175 students, teachers, and leaders about their experiences using them.
The field needs more contributions like these. But neither is built to quantitatively answer the question, “If I buy this, will my students learn more?”
Answering that requires comparison groups, outcomes data, and an implementation window. Accelerate selected providers already integrated into schools, and paired them with an independent research partner to measure how their products affected learning. We funded studies with quasi-experimental designs built to meet ESSA Tier 2 standards, comparing students who used a tool against similar students who did not.
We’ve already seen some interesting stories. At a school in East Palo Alto where Immerse was being used with newcomer English learners, an operations lead noticed that one student in the study group spent her free time logged into the app instead of playing games. In Newark, a KIPP teacher began the year unconvinced. “I was skeptical about using AI in the classroom because I worried it might take away from students’ cognitive development,” she wrote in July. “However, through Quill, I witnessed tremendous growth in my scholars’ ability to read for evidence and incorporate textual evidence into their writing.” In Arizona, Deer Valley administrators went on camera describing the feedback teachers gave about what Goblins did for their middle schools, saying, “We have to have this. My kids cannot be successful to this degree without this available.”
These anecdotes are great, but a tool that students and teachers like isn’t necessarily one that improves learning. That is exactly what our research is designed to evaluate, and we look forward to sharing the findings in an upcoming series of blog posts.
These studies will not definitively settle every question about AI in classrooms, but they represent an important step toward understanding which tools help students learn and under what conditions. We hope they will also inspire education leaders and researchers to join us in identifying what works and scaling approaches proven to deliver results for students.
Jason Godfrey is Managing Director of Data & Information Systems at Accelerate.
