What happened
University of Toronto researchers Philip Oreopoulos and Nina Low followed students in 18 Tennessee middle schools from 2024 to 2026, according to The Hechinger Report. In a draft paper circulated by the National Bureau of Economic Research in August, they found students initially tried Khan Academy’s Khanmigo AI math tutor but largely stopped when it offered hints and questions rather than answers.
Why it matters
The randomized trial tested low-achieving students at least one grade level behind, during an extra remedial math period. Students assigned to Khan Academy with Khanmigo did better in math than peers in regular remediation using tools such as Waggle, IXL, Zearn and DeltaMath, but the gains were modest and no greater than previous Khan Academy practice studies without an AI assistant. For educators and designers, the practical issue is not only whether a tutor can model Socratic help, but whether learners choose to engage when help requires effort.
What to do next
Treat AI tutoring as an instructional design problem, not a plug-in. If piloting similar tools, track when students request help, abandon help, or use it after errors, alongside assessment outcomes. Sal Khan told Hechinger the tested 2023 version has since been integrated into Khan Academy and now appears after wrong answers; schools can also turn it off. Those changes may affect engagement, but evidence is still needed on whether regular AI-tutor use improves learning.
About this briefing
Reviewed by TutorFlow Editorial. We link the primary source, preserve its publication date, and distinguish reported claims from TutorFlow analysis. Our commentary focuses on practical decisions for educators and training teams.
Discussion (0)
Open to everyone