But it's not all smart, clean, good code... I've seen AI repeatedly make the same kinds of errors and interpretations that I would expect from a human working on something. I find that more time in planning, (pre)documentation and testing, even some TDD helps a lot.

I agree, that AI generated code will really start to piss in the pool so to speak. I'm not sure the models will get better without a lot of hand curation and signals of what is good vs bad vs popular code. They emphatically are not the same.