Harvard and MIT Randomized Trial Finds AI Tutor More Than Doubles Learning Gains of Active Learning Class

Harvard and MIT researchers ran a randomized controlled trial in which 194 college students, over two consecutive weeks, experienced both a 60-minute peer active learning class and an at-home lesson delivered by a custom AI tutor called PS2 Pal, covering surface tension and fluid flow. Students using the AI tutor showed median learning gains more than double those in the active learning classroom and reported higher engagement and motivation. The study says PS2 Pal was built around verified content, structured practice, timely feedback and self-paced learning.

The study moves AI tutoring from tool discussion to learning-outcome comparison: with 194 college students over two weeks and two lessons, median learning gains with the AI tutor were more than double those of the active learning classroom, with higher reported engagement and motivation. For schools and teachers, this shifts evaluation criteria toward instructional design quality; for AI education product teams, verified content, structured practice, timely feedback and self-paced learning become verifiable design requirements. Limitations include a college sample, a two-week period and only two physics topics, so results cannot be extrapolated to K-12 or long-term academic performance. Schools and teachers evaluating AI tutoring products will focus more on verified content, structured practice, timely feedback and self-paced design rather than model capability alone. AI education developers gain a comparable instructional design benchmark, making content verification and feedback mechanisms a core area of product differentiation.