AI in Education AI Minute Newsroom 2026-08-20

Two school years, 18 schools, one AI tutor: the gain was about what practising without it produces

Philip Oreopoulos of the University of Toronto and Nina Low of Charles River Associates published a randomised trial of Khan Academy's Khanmigo tutor as an NBER working paper on 17 August. It ran across 18 middle schools in Hamilton County, Tennessee, over the 2024-25 and 2025-26 school years, covering 6,902 student-term observations in 53 grade-level clusters. The pooled effect was 1.26 national percentile ranks per term, with a standard error of 0.60. The second year was stronger than the first, 0.084 standard deviations against 0.020, and the first-year result was not statistically significant. The authors are careful about what that buys. The gain, they write, resembles what Khan Academy practice produces without AI assistance, and because the tutor shipped as part of an integrated program the study cannot isolate what the tutor itself contributed. The chat transcripts explain much of it: only 14.5 percent of student messages contained a mathematical question or any reasoning, while 39.4 percent were bare answers and 24.2 percent were clicks on a suggested prompt.

Why it mattersThis is the evidence that has been missing from the argument about AI in classrooms — not a pilot, not a satisfaction survey, but two school years with a control group. The result is small and real, and the authors point straight at the thing districts keep skipping: the binding constraint was not how capable the model is, it was whether a thirteen-year-old types anything worth answering. Three-quarters of the messages were not questions. Any school weighing a licence should read that number first, because it decides whether the rest of the product matters at all, and it is not fixed by a better model.

✓ Verified · 2 sources

Two school years, 18 schools, one AI tutor: the gain was about what practising without it produces

Related stories