Harvard study: AI tutor doubled learning gains compared with active learning classes
By Greg Kestin, Kelly Miller m.fl. (Harvard University)Built and shared by Greg Kestin, Kelly Miller m.fl. (Harvard University), not by me. Summarised from the public sources below. Check them before you rely on a number.
- Who
- Greg Kestin, Kelly Miller m.fl. (Harvard University)
- Category
- Education
- Source
- Web
- Dates
- Published 3 Jun 2025 · Found 17 Oct 2026
Physicists at Harvard built an AI tutor based on the same pedagogical principles as their active learning classes and tested it in a randomized trial. Students learned significantly more in less time with the AI tutor.
What they built
A custom-built tutor based on GPT-4 for the introductory physics course PS2. System prompts and exercises were designed from research-based pedagogy with scaffolding, feedback and self-pacing. In a randomized crossover trial with 194 students, the tutor was compared with an active learning class session.
The result
Median learning gains were more than twice as large with the AI tutor (post-score 4.5 vs. 3.5), with an effect size of roughly 0.73–1.3 standard deviations. Median time was 49 minutes, and students felt more engaged and motivated.
Sources
Danish upper secondary schools give students an AI tutor that teachers control
By Rosborg Gymnasium & HF and Learnmark Horsens (with Miira)
Rosborg Gymnasium & HF in Vejle and Learnmark Horsens use the Danish AI tutor Miira to help students with written assignments. Teachers decide how the AI may help, and students' conversations with it can be shared with the teacher.
Aarhus BSS experiment: AI lifts weak students in exams but drags down the strongest
By Carsten Bergenholtz, Oana Vuculescu, Franziska Günzel-Jensen and Lars Frederiksen (Aarhus BSS, Aarhus University)
Four researchers at Aarhus BSS tested what happens when business students get a generative AI chatbot in a time-pressured case exam with no clear answer. Weak students improved and strong students got worse, so the grades converged and the exam stopped showing real differences in skill.
Want this in
your business?
Tell me which case caught your eye. I will tell you straight whether it fits, and what it would take.
Start a conversation →