96% of college students tried the AI tutor at least once, but the median student only wrote to him one day in three. The gain looks like that of Khan Academy without AI.
On August 17, 2026, Philip Oreopoulos and Nina Low released the first large-scale experimental results on generative AI in the NBER series. The experiment ran two school years, 2024-2025 and 2025-2026, in 18 colleges in Hamilton County, Tennessee, with random assignment to Khan Academy equipped with Khanmigo tutor, set to guide rather than deliver the answers.
The measured effect is positive but modest: 1.3 percentile national mathematical gain rank per quarter, i.e. 0.06 to 0.08 standard deviations over a school year, and up to 0.14 standard deviations for a year of actually active participation. The authors note that this order of magnitude is similar to that obtained with Khan Academy exercises without IA assistance.
The explanation is less about the capacity of the model than about actual use. If 96% of the students asked for Khanmigo at least once, the median student wrote to him only one third of the days he was training, and only engaged in an exchange in 17% of the sessions where he was making a mistake — when a tutor would have something to do.
Read the full article on the original source
Go to the source →