Every team I talk to measures how much faster they ship now. Almost none of them measure how much longer it takes to debug something they didn't write, because an AI wrote it, twenty minutes ago, and the engineer reviewing it skimmed rather than read. Speed of generation and speed of comprehension used to move together; you couldn't type code faster than you could think it through. AI broke that link. The result isn't bad code, exactly, it's code nobody in the room has fully loaded into their head, which is a much more expensive kind of debt to carry. The fix isn't slowing down generation. It's treating comprehension as a deliverable in its own right, not a side effect of writing the thing yourself.