A tutor for every learner: the two-sigma dream and honest AI
In 1984 Benjamin Bloom showed that one-to-one tutoring lifts the average student above almost all of a regular class. Here is how we think AI can help reach that, and where it must not pretend.
In 1984 the educational psychologist Benjamin Bloom published a short paper with a striking result. Students taught one-to-one by a good tutor performed about two standard deviations better than students taught in a conventional class: the average tutored student did better than roughly 98 percent of the classroom group. Bloom called it the two-sigma problem. The problem was not the result; it was the cost. No society can give every child a personal tutor.
Why tutoring works
A tutor notices what you misunderstood a moment ago. A tutor asks you to explain, waits for your answer and corrects it. A tutor moves on when you are ready and goes back when you are not. In other words, a tutor does by instinct what the research on learning recommends: frequent retrieval, immediate and specific feedback, and practice paced to the learner.
Guidance matters, especially at the start
It is tempting to think that the best learning is pure discovery: give learners resources and let them find their way. For beginners the evidence says otherwise. Paul Kirschner, John Sweller and Richard Clark (2006) reviewed the research and concluded that minimal guidance is less effective and less efficient than guided instruction for novices, because new learners do not yet have the knowledge to direct their own search. A good tutor guides first and steps back later.
What AI can do
For the first time, part of what a tutor does can be offered to everyone. On Universal School, every lesson has an AI tutor that answers questions about that lesson, quizzes at three levels that explain each answer, and a daily review that brings back what you are about to forget. Teachers keep the important role: they set the course, choose what is required, review workshops and see where learners struggle.
Where AI must not pretend
An AI tutor that sounds confident and is wrong does real harm. So we hold ours to three rules. It answers from the lesson and from real sources. Every research reference it gives comes from open scholarly databases, never from the model's memory. And when it does not know, it says so. That is less impressive than a machine that always has an answer, and much more useful.
The goal
We will not solve Bloom's problem in one release. But every learner who gets a patient explanation at the moment of confusion, a question that makes them think, and a reminder before they forget, is a little closer to having the tutor Bloom described. That is the school we are building.
References
- Bloom, B. S. (1984). The 2 sigma problem: The search for methods of group instruction as effective as one-to-one tutoring. Educational Researcher. https://doi.org/10.3102/0013189X013006004
- Kirschner, P. A., Sweller, J., & Clark, R. E. (2006). Why minimal guidance during instruction does not work. Educational Psychologist. https://doi.org/10.1207/s15326985ep4102_1
- Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research. https://doi.org/10.3102/003465430298487