Why Teaching Works
For centuries, the best way to truly understand something has been to teach it. Protegz turns this into a scalable learning system — by giving every student an AI protégé to teach.
The Protégé Effect
The Protégé Effect, first formalized by Jean-Pol Martin in the 1980s, describes a striking phenomenon: people learn more deeply when they teach others than when they study for themselves.
When you prepare to teach, your brain shifts gears. You organize information more carefully, identify gaps in your own understanding, and build stronger mental models. You move from passive recognition to active retrieval and explanation — the deepest forms of learning.
From Betty’s Brain to AI Protégés
In 2009, researchers at Stanford and Vanderbilt created Betty’s Brain — an animated character that middle school students could teach about science. The results were remarkable: students who taught Betty spent more time with the material, engaged more deeply, and outperformed those who studied for themselves.
The key insight was emotional investment. Students experienced genuine frustration when Betty got questions wrong and real pride when she succeeded. They weren’t just memorizing facts — they were building understanding so that someone else could use it.
But Betty’s Brain was a scripted simulation. It couldn’t truly learn, couldn’t ask surprising questions, couldn’t genuinely misunderstand in the way a real student does.
ALTER-Math: Large-Scale Proof
In 2024, the ALTER-Math study brought the teachable agent approach into the modern AI era with a rigorous experiment involving over 6,000 students. Students who taught an AI agent math concepts showed 1.56× learning gains compared to a control group — a statistically significant and practically meaningful improvement.
The study found that the teaching approach was particularly effective for students with lower prior knowledge. Rather than widening achievement gaps, teaching an AI partner helped close them — students who started behind made the largest gains.
Critically, ALTER-Math demonstrated that modern LLMs can serve as genuinely responsive learning partners. Unlike Betty’s Brain’s scripted responses, today’s AI can ask surprising follow-up questions, make plausible mistakes, and genuinely misunderstand in ways that force deeper explanation from the student.
What Protegz Does Differently
Today’s AI is being deployed as a teacher — answering questions, generating explanations, solving problems for students. This is the passive consumption model. Students scroll, consume, and move on.
Protegz flips this entirely. Our AI plays the role of a student — a protégé that your child teaches. The AI genuinely uses only the strategies and explanations your child provides. When it takes a test, it can only apply what it was taught.
This creates a powerful feedback loop: if the AI fails a problem, your child knows their teaching wasn’t clear enough. If it succeeds, they know they’ve truly mastered the concept well enough to transfer it to someone else.
The result is something no AI tutor can achieve: learning driven by the student’s own effort to explain, not by the AI’s ability to explain.
The Research
Students using Betty's Brain, a teachable AI agent, spent more time with material and learned it more thoroughly than those studying for themselves.
Stanford & Vanderbilt, 2009
Students who actively taught material significantly outperformed those who only prepared to teach.
Applied Cognitive Psychology, 2013
Learners preparing to teach employed 1.3x more metacognitive strategies — organizing, evaluating, and connecting ideas.
Memory & Cognition, 2016
In a study of over 6,000 students, those who taught an AI agent (ALTER-Math) showed 1.56x learning gains compared to a control group. The teachable agent approach was particularly effective for students who started with lower prior knowledge.
ALTER-Math Study, 2024
Venkat Narayanan has spent over a decade working in AI and robotics — from earning a Ph.D. at Carnegie Mellon to building and launching America’s first commercial autonomous semi-trucks at Aurora. Along the way, he mentored students from high school through graduate school, and guided engineering teams through hard technical problems.
Through all of it, two things stayed with him. His own understanding grew the most when he was the one explaining — preparing to teach a concept exposed gaps he didn’t know he had. And there was a deep joy in the process, in learning something hard and then helping someone else understand it too.
Protegz is built on these observations. It gives every student an AI protégé to teach — one that’s patient enough to listen, and just fallible enough to make the kind of mistakes that demand real understanding to correct. Understanding that lasts — not just long enough for the next test.