AI Tutoring with Khanmigo in a Two-Year School Experiment

August 2026

Generative AI has been promoted as the innovation that could change learning by providing all pupil a individual tutor. We provision several of the archetypal large-scale experimental evidence, from a two-year collection randomized test in 18 Tennessee center schools in which randomly assigned students used Khan Academy alongside its AI tutor, Khanmigo, configured to trainer fairly than provision answers, during existing regular remedial math sessions. Assignment raises math accomplishment by 1.3 national percentile ranks per term, or concerning 0.06 to 0.08 norm deviations complete a academy year; the implied consequence of a complete twelvemonth of energetic involvement reaches 0.14 norm deviations. These gains resemble those from Khan Academy custom without AI assistance. One clarification is that students used the tutor infrequently and, whenever they did, rarely busy it in substantive mathematical dialogue: 96 percent of students tried Khanmigo at smallest once, but the median pupil messaged it on lone a third of the days they practiced, and in lone 17 percent of the exercise sessions in which they made a mistake. Messages that students did dispatch were mostly naked answers or clicks on suggested prompts. The binding constraint appears to be engagement: realizing the commitment of AI tutoring volition necessitate getting students to use it, not fair giving them access.