A NBER experiment on more than 6,000 middle school students shows that the help of AI slows down progress but improves recovery after mistakes — provided that it is coupled with a logic of mastery.
Philip Oreopoulos, Michael Liut, Alp Sungu and Nina Low conducted a randomized field experiment with more than 6,000 middle school students in Hamilton County schools for the National Bureau of Economic Research, on the NUMI assisted learning platform. The results, published in August 2026, contrast with the usual discourse on automated tutoring.
AI-supported students answer more accurately when they reach an attempt, but progress more slowly and deal with fewer questions overall. The time spent on each question with structured support increases. Above all, the system improves the accuracy of the next attempt and reduces the number of attempts needed to return to a correct answer: the effect is best read on the error retake.
Delayed learning gains, on the other hand, only become marginally significant when AI is combined with progression through mastery, especially on content that has already been worked on. The authors' lesson is clear: a language model only improves tutoring if it is inserted into a device that transforms the mistake into an opportunity to learn, and not into a shortcut to the answer.
Read the full article on the original source
Go to the source →