Harvard tested an AI tutor on 194 students. What happened to their learning?

A Harvard Gazette report revealed striking preliminary findings from a Harvard physics experiment involving 194 students. Those using a custom AI tutor showed about twice the learning gains of students in an active classroom setting, while also reporting greater engagement and motivation. The tutor was carefully designed by faculty. A September 2024 Harvard Gazette report revealed striking preliminary findings from a Harvard physics experiment involving 194 students. New Delhi,Sep 29, 2026 11:45 ISTWhat if your physics tutor could answer questions at any hour, adapt to your pace and never lose patience?That was the idea behind a Harvard experiment that tested a custom-built AI tutor with nearly 200 students. The preliminary results were striking: students using the AI tutor showed learning gains that were about twice those of students in an instructor-led active learning setting.The findings were reported in September 2024 by the Harvard Gazette, based on preliminary analysis of a study led by Harvard lecturers Gregory Kestin and Kelly Miller. The final research was still pending publication at the time.STUDENTS LEARNED MORE, AND ENJOYED IT MOREThe study involved 194 students taking Physical Sciences 2, a Harvard course for life sciences majors. Students experienced two lessons over consecutive weeks. One group first took part in an instructor-guided active learning lesson, while the other worked with the AI tutor at home. The groups then switched approaches the following week.Researchers used pre-tests and post-tests to measure content mastery, while also asking students about their engagement, enjoyment, motivation and growth mindset. According to the preliminary analysis, learning gains in the AI-tutored group were about double those in the classroom group.The researchers also found that students reported greater engagement and motivation while using the AI tutor.The difference was particularly interesting because the comparison was not simply between AI and a traditional lecture. Harvard's classroom model already used active learning and research-based teaching methods.“It was shocking, and super exciting,” Miller said in the Harvard Gazette report.THIS WAS NOT JUST CHATGPTThe Harvard experiment did not simply hand students a blank ChatGPT window.Kestin and Miller designed detailed instructions for the AI to behave like an experienced instructor. The system used a framework built on the GPT application programming interface, with pre-vetted conversations, content-rich prompts and structured feedback.The researchers believed two features helped: personalised feedback and self-paced learning.A student already familiar with a concept could move faster, while someone struggling could spend more time asking questions and working through the material.That could also change what happens inside a classroom.Instead of using class time for students' first exposure to basic concepts, teachers could potentially use it for advanced problem-solving, projects and group work.HARVARD DID NOT SEE AI AS A TEACHER REPLACEMENTThe researchers were also clear about the limits.Kestin warned that AI could either strengthen learning or undermine it depending on how it was used. The goal, he said, should be to help students develop critical thinking rather than have AI “think” for them.Harvard colleagues were already adapting the approach for other subjects. A version was planned for a multivariable calculus course, while Harvard's Derek Bok Center for Teaching and Learning and Harvard University Information Technology were preparing pilots in other large introductory courses.The bigger lesson from the 2024 experiment was not simply that AI can answer students' questions.It was that an AI tutor designed around a teacher's expertise, structured feedback and learning goals could potentially make students more active participants in learning.- Ends