BIGGER PICTURE Learning & work
The physics tutor that waits for your next step
A lecture moves at one speed. A carefully designed AI tutor can stop at the exact step a student has not understood—and ask another question before revealing the answer.

- Study
- 194 Harvard physics students
- Comparison
- More than twice the immediate learning gain
- Scope
- Two lessons; short-term assessments
How the idea works
Choose the next physics step
Vetted course content and software define the progression.
Make the student reason
Questions and feedback respond to the learner’s attempt.
Check the learning
Independent short tests compare gains across teaching conditions.
A private pause button for difficult physics
Harvard physicist Greg Kestin wanted students to practise reasoning without waiting for an instructor to reach their desk. His course-specific tutor broke problems into manageable steps, offered feedback and let learners set the pace. A confident student could proceed; someone struggling could stay with one idea until it made sense.
Sources: [2]
The teaching design did the heavy lifting
The researchers did not simply open a chatbot and ask it to teach physics. They supplied vetted course content and designed the interaction around active thinking, cognitive load and encouragement. When prompting alone failed to control the sequence reliably, the surrounding software enforced the intended progression. The educational structure was part of the intervention.
Sources: [1]
Two lessons, a striking comparison
In a randomized crossover experiment, 194 eligible students encountered both AI-supported learning and classroom active learning. Immediate pre- and post-tests showed more than twice the learning gain with the tutor, with less time spent. Students also reported stronger engagement. This is evidence about a particular tutor and carefully selected lessons, not every AI conversation.
Sources: [1]
Preparation can make classroom time more valuable
Kestin described the tool as preparation for richer classroom work. That distinction matters: immediate test performance does not establish lasting understanding, semester grades or the value of replacing teachers. The practical opportunity is narrower and useful—give students responsive practice before they arrive, then use human teaching for discussion, judgement and the misconceptions that remain.
Sources: [2]
The original sources
- Kestin and colleaguesAI tutoring outperforms active learning: author manuscript
- Harvard GazetteProfessor tailored AI tutor to physics course. Engagement doubled.