A free preview, generated by MasteryLoop HQ.
Harvard physics researchers ran an actual controlled trial: 194 students in an intro physics course, split so every student experienced both conditions in a crossover design — one week of the school's own hands-on active-learning classroom, one week with an AI tutor at home. Same course, same material, same pedagogical design, only the delivery changed.
Students using the AI tutor scored roughly double the learning gains of the in-class group on the post-test, with a large effect size (0.73 to 1.3 standard deviations, from quantile regression). They also reported higher engagement and motivation. Median time on task: under an hour.
The tutor didn't hand out answers. It asked questions, pushed back, and wouldn't let a student move on without doing the thinking themselves. Researchers credit that structure — not the AI itself — for the result.
This was one course, two topics — surface tension and fluid flow — not a referendum on higher education. The authors say plainly they don't expect this to hold for material requiring complex synthesis of many ideas or real critical thinking.
Most people's daily experience of 'AI learning' is a chat window that answers on request. This study didn't test that. It tested a system that gates progress behind proof of understanding — a structural design choice, not a personality trait of the AI.
The interesting finding isn't 'AI is good at teaching.' It's that how the AI is built to behave — forcing engagement instead of allowing passivity — is what produced the gap. That's an argument about design, not about AI in general.
Think about the last time you used AI to learn something. Did it check whether you actually understood it, or did it just answer and move on?
Reading a good explanation feels like understanding it. Usually it isn't the same thing — and you don't find out which one you've got until someone asks you to explain it back.