A growing but still narrow slice of U.S. school districts are moving from writing AI policies to putting AI tools in front of students and teachers. The 2025-26 school year has brought vendor contracts and classroom pilots in districts spanning Pennsylvania, New Jersey, Florida, and New York, even as the most rigorous research on whether these tools help students learn remains drawn from settings outside American K-12 classrooms.
In June 2026, the Haverford Township School District board in Pennsylvania voted 5-3 to approve one-year contracts with School AI at $12,999 for student-facing AI tutors and Brisk at $22,260 for teacher task automation. Administrators told the board that 97 teachers had already been using a free version of Brisk, according to The Philadelphia Inquirer. The same report noted that administrators said Brisk would not grade assignments but would provide what they called first-level feedback, such as flagging a missing topic sentence, with teachers then reviewing and revising that feedback in their own voice. School AI was scoped for high school and potentially middle school, not elementary.
Parent critics at Haverford argued that students would inevitably use AI tools inappropriately, making it easier to avoid work. Administrators countered that the tools would not replace teaching and learning. That tension between deployment and trust is playing out as the national footprint of classroom AI remains small.
How wide is deployment, really
Bellwork's 2026 readiness report found that about 22.6 percent of the 18,301 U.S. school districts show some public AI activity, but only 13.0 percent have reached what the report calls guided use or implementation stages. Just 5.8 percent show visible day-to-day implementation. Twelve point eight percent have published an AI policy.
ChatGPT is the dominant tool by a wide margin, appearing in 3,121 districts, according to the same report. Other deployments take different forms. Newark Public Schools in New Jersey demonstrated AI-powered learning alongside AFT President Randi Weingarten, signaling union-level engagement. Flagler County Schools in Florida reported rising AI integration in classrooms. Salamanca City School District in New York announced a pilot of an AI robot teaching assistant, making it one of the first U.S. districts to test a physical AI-robot classroom assistant.
What the causal evidence says, and where it comes from
A 2026 Stanford-led review of more than 800 academic papers on AI in K-12 found that only 20 produced strong causal evidence. The review also found that no high-quality causal studies of student-facing AI tools had been conducted in U.S. K-12 school settings. Existing evidence comes from international settings, students above age 18, and short-term lab experiments, limiting how cleanly it applies to American classrooms.
On the question of whether AI tools help students learn, the findings are mixed. The Stanford review found that AI tools significantly improve student performance on math, programming, and writing tasks while students have active access to them. But when tools are removed, effects are mixed, with some experiments showing worse learning outcomes from general-purpose AI chatbots.
The same review found an established pattern that tools designed with pedagogical guardrails show more promise than general-purpose AI. Tutoring chatbots that provide step-by-step reasoning rather than direct answers appear better positioned to support durable learning.
What international experiments show
Several international studies cited in the Stanford review illustrate the range of outcomes. An experiment in Turkey found that high school students who practiced for an exam using a general-purpose AI chatbot performed worse on the final exam, while those using a tutoring-specific AI chatbot performed the same as peers without AI, suggesting that tool design mediates learning transfer. An experiment with college students in Germany found that students using general-purpose AI chatbots to conduct research demonstrated lower-quality reasoning and argumentation compared to those using a traditional search engine, raising concerns about cognitive offloading. A study in Brazil, by contrast, found that giving high school students AI-generated feedback on their essays improved their scores on a high-stakes argumentative writing exam, offering a positive precedent for automated feedback in writing.
Teacher-side uses and the implementation gap
The Stanford review also examined teacher-facing applications. Teachers using AI tools for lesson preparation spent less time on planning without reducing lesson quality, and AI tools providing regular automated feedback to human tutors improved instructional quality, especially for less experienced and lower-rated tutors. Those findings are characterized as mixed in the review.
Access alone does not guarantee engagement. A study reported by The 74 Million found that simply giving students access to AI tutors did not mean they would use them, highlighting an implementation gap between tool availability and actual student engagement.
The U.S. Department of Education's AI report draws a parallel to prior education-technology waves, including one-to-one laptop programs, which showed that simply providing devices did not reliably improve academic outcomes. Benefits depended on pedagogical integration and teacher support, a pattern researchers note may recur with AI tools.
The distance between the 22.6 percent of districts showing some public AI activity and the 5.8 percent showing visible day-to-day implementation tracks with the gap researchers have flagged between tool availability and tool use. Districts like Haverford are betting that vendor contracts and board votes will bridge that distance. The evidence on whether that bet pays off for students is still being built, largely outside the classrooms where the tools are now arriving.