A new policy framework from the Manhattan Institute calls for strict grade-level restrictions on artificial intelligence use in K-12 classrooms, proposing that students in kindergarten through second grade have no direct interaction with AI tools at all. The report, published in May 2026, argues that generative AI's ability to complete academic work on behalf of students poses a unique threat to learning that differs fundamentally from earlier classroom technologies like film, computers, or tablets. Unlike those tools, which changed how information was presented but still required students to do their own thinking, AI can write essays, solve math problems, and perform the cognitive tasks that learning depends on.
The framework organizes AI policy around students' developmental readiness rather than the technology's capabilities. In grades K-2, AI would be restricted to teacher-facing uses like lesson planning and assessment analysis, with no student interaction permitted. Students in grades 3-5 would engage with AI only under direct teacher supervision, completing all assignments independently before analyzing AI-generated responses as objects of critique. Middle schoolers in grades 6-8 would be required to produce work first and use AI only afterward for comparison, with a formal AI literacy curriculum made mandatory. High school students would gain expanded but still constrained access, using AI to identify gaps in their reasoning or find counterarguments only after completing independent work. No summative assessments at any grade level would allow AI use, except possibly in dedicated AI literacy courses.
The report warns that current adoption patterns mirror historical mistakes with educational technology. By 2024, the average K-12 district was using nearly 2,700 different educational technology tools annually, more than triple the number before the pandemic, yet most of those tools entered classrooms without evidence they improved student outcomes. The authors cite a 2025 federal lawsuit alleging that i-Ready, one of the most widely used diagnostic platforms, collected and shared sensitive student data with third parties for years without parental consent. They also point to Los Angeles Unified's $1.3 billion iPad initiative in 2013, which produced no measurable improvement in learning and ended in fraud investigations. According to the report, research on AI tutoring has found that students given unrestricted access to large language models performed worse on later assessments than control groups, while even constrained AI tutors showed neither harm nor benefit.
The authors argue that learning requires cognitive effort, the mental struggle to retrieve information, reason through problems, and work through difficulty. When AI completes a task instead of the student, that effort disappears, a process researchers call "cognitive offloading." The risk is especially acute for younger students building foundational skills: research shows that phonemic awareness assessed in kindergarten predicts reading achievement in third and fourth grade more reliably than IQ or family income. Early number sense in kindergarten predicts fifth-grade math achievement. Students who don't master these basics by upper elementary school face compounding disadvantages in later grades. In 2024, only 31% of New York fourth-graders scored at or above proficiency in reading, a 9-point drop from a decade earlier. The report contends that AI use without developmental guardrails will widen those gaps, particularly for Black, Latino, and low-income students who are already most likely to lack foundational skills.
The framework recommends that New York establish enforceable statewide guidelines governing student AI interaction, strengthen early literacy and numeracy instruction, and create a public registry of approved AI platforms that meet developmental, privacy, and efficacy standards. Districts would be required to publicly disclose all AI tools in use, including what student data they collect and whether that data trains commercial models. A voluntary statewide procurement consortium would allow districts to access pre-negotiated vendor contracts with standardized protections, redistributing capacity so that smaller or poorer districts aren't left negotiating alone. The report also calls for state guidance to help families limit AI use at home, noting that a December 2025 survey found 62% of middle school students already using AI for homework. The bottom line: without grade-level protections rooted in how children actually learn, schools risk raising a generation that can prompt AI but can't think independently.