The research behind C1
C1 is a study coach that teaches by talking with the learner: it asks questions instead of handing over answers, and draws on a whiteboard while it speaks. Each of those choices follows published research. This page lists that research, what we built from it, and what it does not tell you.
What the research found, and what C1 does with it
| Finding | The evidence | In C1 |
|---|---|---|
| Asking, not telling | In a field experiment with about 1,000 high-school maths students in Turkey, students given plain ChatGPT did better while they had it, but 17% worse on the exam once it was taken away than students who never had it. A tutor built to give hints instead of answers largely removed that harm.1 Earlier dialogue tutors that teach through questions, such as AutoTutor, averaged learning gains of about 0.8 standard deviations.2 | The tutor never says the answer. It asks, gives a hint at a time, and the answer appears on the board only when the learner finds it. |
| A planned lesson, not an open chat | In a randomised trial in Harvard's largest physics course (194 students), students learned more than twice as much from an AI tutor as from an active-learning class, in less time. The authors credit the teaching built into the tutor: it set the order of the steps, worked from prepared step-by-step solutions and kept the load manageable.3 | Every lesson follows a board plan of short steps, prepared before the lesson; the tutor teaches along it instead of improvising. |
| Words and pictures together | People learn better when the matching words and pictures come at the same time rather than one after the other (the temporal contiguity principle).4 | The board draws the step the tutor is talking about, as it talks: a fraction, a balance, a cell, the water cycle. |
| Point at what matters | Cues that highlight the essential part improve learning (the signalling principle: supported in 24 of 28 experiments, median effect size 0.41).4 | A teacher's pointer moves to the thing being talked about, or to the "?" being asked. |
| Regular use is what counts | In a two-year trial in 18 Tennessee middle schools, an AI tutor that coached rather than answered raised maths scores only slightly (about 0.06–0.08 standard deviations a year). The limit was use: most students rarely asked it for help.5 | Plans counted in short lessons of 15 minutes (20 a month), a weekly progress report to parents, and parental controls for a daily time limit and bedtime, so studying becomes a small daily habit. |
AI tutoring can work at scale
A six-week after-school programme with a GPT-4 tutor in Nigeria (759 students, randomised) raised results by 0.31 standard deviations, which the researchers compare to 1.5–2 years of typical schooling.6 Google reports that in a randomised trial with 1,763 students in Sierra Leone, at least 12 hours with its Guided Learning tutor over eight weeks moved a typical student from the 50th to the 64th percentile in maths.7 Both programmes were guided towards learning, not answer-giving.
What these studies do not tell you
They are studies of other tutors, other subjects and other countries. They show what kind of design tends to help; they are not evidence about C1 itself. C1 has not yet been evaluated in an independent study.
What we do measure: before a new topic the learner can take a short three-question check, and the same kind of check afterwards, so each family sees what changed. When we have enough of these results, we will publish them here, including if they disappoint.
References
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning: evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. Working paper · Summary (Wharton)
- Graesser, A. C. and colleagues, AutoTutor, University of Memphis. Overview and effect sizes
- Kestin, G., Miller, K., Klales, A., Milbourne, T. and 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, 15. Article (open access)
- Mayer, R. E. (2020). Multimedia Learning (3rd ed.). Cambridge University Press: temporal contiguity and signalling principles. Chapter
- One Click Away: AI Tutoring with Khanmigo in a Two-Year School Experiment (2026). NBER Working Paper 35620. Paper
- World Bank (2025). From chalkboards to chatbots: transforming learning in Nigeria, one prompt at a time. Report
- Google, LearnLM and Gemini Guided Learning: results reported by Google. LearnLM