Research
Why we build ClassDay as structured teaching.
Our design choices are informed by research on what makes AI tutoring actually work, not the assumption that any chatbot will do.
The study
Kestin, G., Miller, K., Klales, A., Milbourne, T., & Ponti, G. (2025). AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting. Scientific Reports. Read the paper on nature.com
What it did
A randomized controlled crossover study in an introductory Harvard physics course (N=194) compared a carefully designed AI tutor against in-class active learning, using the same lesson content for both conditions.
What it found
- Students using the AI tutor showed median learning gains more than double those in the in-class active-learning condition for the tested lessons.
- Median time on task was 49 minutes for the AI group versus 60 minutes in class.
- Students reported significantly higher engagement and motivation with the AI tutor.
- 83% of students reported the AI tutor's explanations were as good as or better than their human instructors' explanations in that class.
What it didn't show
The authors were explicit that these results depended on a research-based pedagogical design: active learning, scaffolding, managing cognitive load, accuracy, targeted feedback and self-pacing. The study warns that unstructured use of a general chatbot like ChatGPT is not equivalent, and its results should not be assumed to generalize to any AI tool used casually or without that structure.
Research suggests carefully structured AI tutoring can improve learning outcomes; the study does not claim that all AI tutors, or ClassDay specifically, will reproduce these results.
How this relates to ClassDay
ClassDay's teaching loop, explain, check understanding, give specific feedback, adapt, is designed in the spirit of this research. But this was an independent study of a different AI tutoring system in introductory physics. ClassDay has not yet been clinically or academically validated to produce the same results. We cite this work because it shaped our design, not as a claim about ClassDay's own outcomes.