Korva is a scroll of the internet’s best ideas.
Stanford found only 20 solid studies
The review
The March 2026 report is titled The Evidence Base on AI in K-12: A 2026 Review. It went through the research on AI in schools from kindergarten through twelfth grade and counted several hundred published papers, with 20 meeting the bar for high-quality causal studies. More have appeared since March.
What it found
The review's conclusion is that student performance often improves with access to AI tools, but once the tools are removed, results are mixed. Tool design matters, so not all tools are equal, and AI may meaningfully support educators. The evidence base does not show strong evidence that AI harms kids' education and development.
The tutoring case
The case for the tools rests on personalized tutoring: people learn at different paces and in different ways, visually, orally, through stories, or through interactive project-based work, and each of those can be tailored to the individual.
What makes a study causal?
A causal study can show that one thing produced an effect, rather than merely appearing alongside it. The strongest design is a randomized controlled trial, where chance decides who gets the intervention. Quasi-experimental designs, such as comparing similar schools before and after a rollout, can also qualify. Studies that compare students who chose a tool with those who did not mix up the effect of the tool with the kind of student who picks it up.
Why do gains vanish when tools are removed?
Psychologists Nicholas Soderstrom and Robert Bjork reviewed decades of research in a 2015 paper and concluded that performance during practice is an unreliable guide to learning. Conditions that boost immediate performance, such as help or hints, often do not produce lasting learning. Conditions that make practice harder, such as spacing and testing, often produce better retention despite worse immediate scores. Bjork calls these desirable difficulties.
Why are there so few strong AI studies?
Randomized trials in schools need approvals from districts and parents, and results take at least a full academic year to measure. Peer review and publication add months more. ChatGPT launched in November 2022, so most classroom research on generative AI has had only a few school years to run.
More from All-In with Chamath, Jason, Sacks & Friedberg
