Can an AI tutor help students learn complex subjects more effectively than a university classroom? A Harvard experiment involving 194 undergraduates has offered evidence that artificial intelligence could change how students are taught, particularly in subjects such as physics.The study compared a specially designed AI tutor with an active-learning classroom at Harvard, where students learn through group work and instructor support. Students using the AI tutor recorded more than twice the learning gains of those attending the classroom sessions, while also spending less time on the lessons. The findings, published in Scientific Reports in 2025, have raised questions about personalised learning and the role of technology in higher education.
What happened when Harvard students learned physics with AI?
The research was led by Gregory Kestin and Kelly Miller, who examined learning outcomes among 194 undergraduates enrolled in Harvard’s Physical Sciences 2 course, which teaches physics to life sciences majors.The experiment compared two approaches to teaching the same material. One involved an active-learning classroom, with students working through concepts alongside instructors and peers. The other involved a custom-built AI tutor that students used remotely.The researchers used a randomised crossover design. Students experienced both teaching methods across two lessons, covering topics including surface tension and fluid flow, with the order of the methods varied between groups.According to the published research, students using the AI tutor achieved more than twice the learning gains recorded in the active-learning classroom. Their median post-test score was 4.5, compared with 3.5 for the classroom group.The researchers also reported that students completed the AI-supported lessons in less time and reported higher engagement and motivation.
Why the AI tutor performed differently from a regular classroom
The experiment did not involve students simply opening ChatGPT and asking it to explain physics.The Harvard team designed a specialised tutor using established principles of teaching and educational psychology. It was built to guide students through the material, respond to their questions and provide feedback based on their individual understanding.Instead of immediately supplying answers, the tutor encouraged students to work through concepts and engage with the material.This allowed students to learn at their own pace. Someone struggling with a particular concept could ask additional questions, while another student who understood it could move ahead.In a classroom, an instructor must divide attention among multiple students. Even in a well-designed active-learning environment, individual students may have different levels of prior knowledge and require different amounts of explanation.The AI tutor offered a way to provide more individualised feedback during the learning process. The researchers identified personalised instruction and self-paced learning as important factors behind the results.
Does this mean AI can replace university teachers?
The findings do not establish that AI can replace universities or that classroom teaching has become unnecessary.The experiment involved 194 students at one university, focused on two physics lessons and measured learning through immediate assessments. It did not establish whether the same results would occur across other subjects, institutions or longer courses, or whether students would retain the material equally well over an extended period.The comparison was also between a carefully designed AI tutor and an established classroom teaching method, not between an ordinary chatbot and every aspect of university education.University education includes more than learning course material. It also involves practical work, discussion, collaboration, mentorship and interaction with teachers and fellow students.The Harvard research instead points to a possible way of combining AI tutoring with classroom instruction. Students could use AI to develop an initial understanding of difficult concepts before attending classes where instructors focus on discussion, application and areas of difficulty.
What the experiment could mean for students
For students, the study highlights the potential of AI to make individualised academic support more accessible.A student who struggles with a physics concept may be able to ask questions repeatedly without worrying about holding up a class. Someone preparing for an examination could use a tutor to identify gaps in understanding and practise at a suitable pace.For colleges and educators, the question is how such tools can be integrated into teaching while preserving the aspects of education that require human interaction and guidance.The Harvard experiment provides evidence that a carefully designed AI tutor can improve learning outcomes in a specific undergraduate physics setting. Whether similar approaches can deliver comparable benefits across higher education remains a subject for further research.Disclaimer: This article is based on research published in Scientific Reports and reporting by Harvard Gazette. TOI Education does not independently verify the research findings or conclusions.
