Hamilton County Schools (Khan Academy Khanmigo)
Tennessee schools test Khan Academy's Khanmigo AI tutor in a two-year trial: small gains, thin use
Across two school years, 18 Tennessee middle schools used Khan Academy with its Khanmigo AI tutor in daily remedial math sessions as part of a randomised trial by independent economists. Maths scores rose a little, about as much as Khan Academy practice without AI, and the authors point to thin use: most students rarely talked to the tutor.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
The problem
Schools struggle to give weak math students the one-to-one help that works best, because human tutoring is expensive and hard to staff. Hamilton County Schools in Tennessee already ran daily remedial math blocks for students below grade level. The question was whether an AI tutor inside those blocks could add learning at low cost.
What they built
Students used Khan Academy practice together with Khanmigo, an AI tutor set up to guide with questions rather than hand over answers. It sat one click away while students worked on each problem. Control grades carried on with ordinary remedial lessons, combining teacher-guided practice with other math software that had no AI tutor.
How they did it
Economists Philip Oreopoulos and Nina Low ran a cluster randomised trial. Within each of 18 middle schools, one or two of the three grades were randomly assigned to the AI version. Results were measured with the district's own MAP benchmark test over five terms across the two school years. The team also analysed chat messages from the second year to see how students used the tutor.
Results
Being assigned to the AI tutor raised maths scores by 1.3 national percentile ranks per term (about 0.06 to 0.08 standard deviations across a school year). The authors say this resembles what Khan Academy practice without AI delivers. Nearly all students (96 percent) tried Khanmigo, but the median student messaged it on just a third of practice days and in 17 percent of sessions where they made a mistake. Treated students did about 30 extra minutes of weekly practice. Khan Academy's founder Sal Khan told Hechinger the findings matched the company's own observations.
What went wrong
Engagement was the weak point. Most messages to the tutor were bare answers or clicks on suggested prompts, and only about one message in seven contained a maths question or a step of reasoning. The authors conclude that getting students to use an AI tutor matters more than giving them access.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Gain in math achievement from being assigned to Khan Academy with Khanmigo | 1.3 national percentile ranks per term1 | |
| The same gain expressed over a school year | about 0.06 to 0.08 standard deviations1 | |
| Implied effect of a full year of active participation | 0.14 standard deviations1 | |
| Students who tried Khanmigo at least once | 96 percent1 | |
| Share of practice days on which the median student messaged Khanmigo | a third1 | |
| Share of exercise sessions with a mistake in which the median student messaged Khanmigo | 17 percent1 |
Lessons for your project
Measure real use, not access. Nearly every student tried the tutor, yet the typical student rarely asked it for help, so licence counts would have overstated the value.
Do not count on students to seek help. Khan Academy itself says students were not engaging much, and it has since redesigned the tool to step in automatically.
Build the AI into protected practice time. Here it ran inside existing daily remedial blocks, which gave students extra practice; the authors think the gains came mainly from that practice.
Compare against the same software without AI. The authors found the gains matched what Khan Academy practice delivers on its own, so test AI against that baseline.
Budget for human help where stakes are high. The authors note each student learns more from human tutoring in a year, at far higher cost.
Caveats
- The results come from an NBER working paper by two independent economists. It has not been peer reviewed. Hamilton County Schools and Khan Academy are named as partners, and the paper credits Khan Academy for platform help and data sharing.
- The sources do not say which language model powered Khanmigo in the trial, so the model is shown as undisclosed. The paper says only that Khan Academy launched Khanmigo in 2023 after early access to GPT-4.
- Hechinger reports that the tested version was released in 2023 and that Khan Academy has since changed how it works, so the findings may not apply to today's product.
- The design cannot separate the AI tutor's own contribution from the structured practice. The authors say the gains are mainly from practice.
- The paper says Khan Academy changed the tutor in January 2026 to open by default shortly after page load, partway through the trial.
- Deployment year is 2024 because the trial began in the 2024-25 school year. The NBER page gives an issue date of August 2026 only, so its source date is shown as 2026-08.
Sources
- One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment
National Bureau of Economic Research (working paper 35620) · 2026-08 · Research paper · nber.org
- Students didn't get answers from Khanmigo. They didn't want its questions, either
The Hechinger Report · 2026-09-21 · News · hechingerreport.org
Summarized in our own words from the sources above. Spotted an error? See corrections.