The easiest AI tutor to build is one that answers questions. It's also nearly worthless for learning. Decades of learning science are clear: people learn by retrieving, struggling, and connecting — not by receiving.
Deliberately unhelpful, helpfully
So our tutor is built to be strategically withholding. Ask it for the answer and it asks you a question back — one calibrated to where your understanding actually broke. It spaces review of earlier material into new sessions because spaced retrieval beats cramming. It lets you be wrong for a moment, because productive struggle is where the learning happens, then steps in before struggle becomes frustration.
Measured by learning, not satisfaction
This creates a product tension: learners rate answer-giving tutors higher in the moment. We optimize for the delayed metric instead — performance on later assessments the tutor has never seen. That number, not the star rating, tells you whether teaching happened.
The uncomfortable truth about education AI is that the helpful-feeling product and the effective product are different products. We chose the second one.