Does AI Tutoring Actually Work? A Harvard Study Says Yes — If It's Built Right
Short answer: yes, but only for a specific kind of AI tutor. A randomized Harvard trial found students learned more than twice as much from a carefully designed AI tutor as from an active-learning classroom, in less time. The tutor won because of how it taught: it asked questions before it gave answers.
What did the Harvard study actually test? Most advice about AI and homework falls into two camps: it will ruin how kids learn, or it's the answer to everything. The Harvard study is more useful than either take because it measured something specific. It didn't test "AI in the classroom" in the abstract. It tested one purpose-built AI tutor against one of the strongest alternatives available, an active-learning physics classroom, the kind of teaching that already beats a plain lecture by a wide margin in education research.
Lecturers Gregory Kestin and Kelly Miller ran the trial in their own course, Physical Sciences 2, in fall 2023. It was a randomized crossover study of 194 undergraduates: each student learned one physics topic in the classroom and a different one from the AI tutor, so every student served as their own comparison. The peer-reviewed paper was published in Scientific Reports in 2025.
What were the results? The AI-tutored lessons produced learning gains more than double those of the classroom lessons, and students got there faster: a median of 49 minutes, against 60 in class. Students also reported feeling more engaged and more motivated.
Why did the AI tutor win? The detail that matters more than the headline number is how the tutor was built. Kestin's design rule, as he put it to the Harvard Gazette, was that an AI tutor shouldn't "think" for students but help them build their own critical thinking. In practice the tutor revealed solutions one step at a time, encouraged students to attempt a problem before it showed the answer, and kept its replies short so students weren't overloaded, according to the Hechinger Report's coverage.
That is a description of good tutoring, with or without AI in the sentence.
Does this mean any chatbot works for homework? No, and this is the part that gets lost when the study is shared. "AI tutor beats classroom" reads, to a lot of parents, like an argument for dropping homework into any chatbot. It is closer to the opposite. The tutor won because of how deliberately it withheld the answer, not because it was fast at producing one. A student who pastes a worksheet into a general-purpose chatbot and copies back whatever comes out is doing something structurally different.
What doesn't the study show? It's worth being honest about the limits. The trial ran in a single course at an elite university, with scheduled, proctored sessions, so it can't tell us how students would use a tutor on their own at home for months. Kestin himself has said the result isn't an argument for replacing human interaction. A tutor like this is a strong complement to teachers and parents, not a substitute for them.
What about kids doing schoolwork in a second language? For a family navigating homework in a language that isn't spoken at home, the gap is wider still. Harvard's tutor never had to leave English: it was English physics, explained in English, to English-speaking students. A child doing Arabic-medium schoolwork in an English-speaking household is solving a harder version of the same problem, understanding the concept and bridging the language it's written in at the same time, without a parent who can always check the page. The same patient, question-first tutoring has to work in two languages, not one.
How do you choose an AI tutor for your child? Whatever the subject or language, ask three questions before picking a tool:
Does it ask before it tells? A tutor that hands over the answer is a shortcut, not a teacher. Does it stay with a stuck idea until it lands? However many different ways that takes. Does it fit what your child is actually studying? A good tutor works from the real page in front of the student. The Harvard study is good news, not a warning. It is real evidence that a well-built AI tutor can outperform even strong classroom teaching. It just isn't evidence for the version most families are using tonight.
Sources Kestin, G., Miller, K., Klales, A., et al. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports, 15, 17458. Harvard Gazette. Professor tailored AI tutor to physics course. Engagement doubled. The Hechinger Report. An AI tutor helped Harvard students learn more physics in less time