Harvard RCT Finds AI Tutor Outperforms Active-Learning Classroom

A Harvard team including Kestin and Miller published a randomized controlled trial in the Nature journal Scientific Reports, comparing the teaching effectiveness of a personalized generative-AI tutor against traditional active-learning classrooms. The study assigned 194 undergraduates to learn the same physics content in two groups: one attended Harvard's top-tier active-learning classroom (small-group collaboration plus instructor support), while the other studied at home with an AI tutor. The AI-tutored group achieved more than twice the learning gains of the elite classroom group, scored about 30% higher on the final assessment, spent less time learning, and reported better engagement and motivation. The paper, published in June 2025, has drawn 210,000+ reads and 202 citations as of this post.

Confirmed

The study is a randomized controlled trial published in Scientific Reports by a Harvard team (Kestin, Miller, et al.)
The sample was 194 undergraduates studying the same physics course
The AI tutor group scored roughly 30% higher on the final assessment than the active-learning group, achieved over twice the overall learning gains, and took less time
The AI tutor was designed to follow the same evidence-based pedagogy best practices as classroom teaching
The paper has received 210,000+ accesses and 202+ citations

Why it matters

This is one of the first high-quality RCTs to directly compare a generative-AI tutor with a top in-person classroom in a real university teaching setting, providing empirical evidence for AI's effectiveness in personalized education
The results suggest AI tutors can outperform traditional elite teaching models in both efficiency and effectiveness, with implications for higher-education course design
Multiple authors including @burkov, @humphreyshi, @CurieuxExplorer, and @alvelda considered the finding a landmark for education, and the study's authors themselves raised questions about the value of classroom teaching based on these results

2026-09-28 ~ 2026-09-29 · 5 related posts

Primary sources

Nature RCT: AI Tutor Beats In-Class Active Learning in Less Time — burkov · 2026-09-28 23:46 (UTC+08:00)
Harvard RCT: AI Tutor Beats In-Class Active Learning in Less Time — humphrey_shi · 2026-09-28 23:35 (UTC+08:00)
Harvard experiment: AI tutor students scored 30% higher than elite classroom peers — alvelda · 2026-09-28 20:53 (UTC+08:00)

[source] Harvard experiment: AI tutor students scored 30% higher than elite classroom peers — alvelda · 2026-09-28
[source] Harvard RCT: AI Tutor Beats In-Class Active Learning in Less Time — humphrey_shi · 2026-09-28
[source] Nature RCT: AI Tutor Beats In-Class Active Learning in Less Time — burkov · 2026-09-28

2 near-duplicate retellings: humphrey_shi · CurieuxExplorer