Why answers alone do not teach
There is a long-standing finding in learning science: the effort of retrieving and reasoning is what builds durable understanding. We look at what this means for how educational AI should be designed.
It is tempting to measure a learning tool by how fast it resolves a question. A faster answer feels like faster progress. Decades of research in cognitive science suggest the opposite can be true: when the path to an answer is made too easy, the learning that should accompany it often does not take place.
The principle is sometimes described as desirable difficulty. Effortful retrieval, the act of working something out rather than reading it off a page, strengthens memory and deepens understanding in a way that passive reception does not. A student who struggles productively toward a solution tends to retain and transfer that knowledge far better than one who is simply told.
This has direct consequences for how we build AI for education. A system optimised to be maximally helpful in the narrow sense, one that always gives the complete answer immediately, can quietly undermine the very learning it is meant to support. The student gets the mark today and loses the understanding tomorrow.
Our approach is to design for productive struggle rather than against it. That means calibrating support carefully: enough to prevent a student from giving up, never so much that the thinking is done for them. It means asking questions, offering hints in graduated steps, and stepping back as soon as a student can carry on alone.
None of this is an argument for making learning harder for its own sake. Frustration without progress teaches nothing. The skill, for a human tutor and for a well-designed AI alike, is in judging the right amount of help at the right moment.
We treat this as an ongoing area of study rather than a solved problem. As learning science advances, so should the tools built on it.
