Harvard study finds AI tutor can speed up physics learning

How the experiment was conducted

A randomized trial at Harvard University compared two methods of teaching physics to college students: in-person classes with active learning and an AI-based tutor. The study, published in Scientific Reports in June 2025, involved 194 undergraduate students from an introductory physics course during the fall of 2023. Participants were randomly assigned to minimize bias.

The study used a crossover design. Each student covered one topic with the active class and another with the AI tutor at home in consecutive weeks, with the order determined by a lottery. The topics chosen were surface tension and fluid flow. Knowledge was assessed through tests administered before and after each session.

Results indicate greater efficiency

Students who used the AI tutor learned more than twice as much content in less time. The median session duration was approximately 49 minutes, while the in-person class required 75 minutes. Additionally, participants reported higher engagement and motivation when interacting with the digital tool.

How the tutor works and its limitations

The tutor, named PS2 Pal, was developed using the GPT-4 model and was instructed not to provide direct answers. In practice, students received a problem, attempted to solve it, and then interacted with the system, which offered hints and questions to guide their reasoning. The professor provided the system with the same feedback notes used in the classroom.

The authors caution that the results do not prove AI-structured teaching is superior to active classroom teaching in all contexts. Tasks requiring synthesis of multiple concepts and critical thinking may present additional challenges. Since the sample consists of Harvard students, the applicability of the findings to other student groups remains uncertain.

Gregory Kestin, one of the authors, states that the goal is not to replace teachers but to optimize classroom time. The aim is to reduce time spent on vocabulary and basic concepts, freeing up space for creative activities and more complex problems.

Perspectives and complementary research

Other recent studies support these findings. A survey by Fast Company Brasil indicates that AI improves learning when used as a tool to support teachers, although there is no evidence it surpasses human teachers. A 2026 report by the Brookings Institution highlights that AI enhances digital tutoring systems by enabling students to communicate with the program in natural language.