But they don't. LLMs can't understand messy code much better than humans can. Maybe a little, but not enough to compensate for the code they create being messy.