The Daily Californian headline is the kind that gets passed around with a sigh: failing grades soar as Berkeley CS professors report more AI usage and weaker math skills. The usual chorus about kids these days, and then everyone moves on.
They shouldn't. The cheating frame is the lazy read. Cheating, if you cheat well, produces inflated grades, not collapsing ones. What the Berkeley signal actually describes is students who used a tool that delivered correct-looking answers all the way through the easy parts of the curriculum, and then hit the part where you have to actually hold a proof in your head. The tool didn't fail them at the exam. It failed them eighteen months earlier, silently, by doing the reps they were supposed to do.
This is the part of the AI productivity story that the vendor decks skip. Output goes up. Underlying capability, the thing you only notice when the scaffolding is removed, goes down. In a coding job that gap can stay hidden for years. In a math-heavy CS sequence it shows up by midterm.
The right question for any organisation rolling out copilots is not how much faster the work gets. It is which muscles are quietly atrophying, and when the bill comes due.