I don't really generate big chunks of code, but in general I found AI outputs to be quite sensible. It knows what the main modules are and uses them, I've not seen any as egregious as making a JSON parser from scratch without a need.
Maybe not a JSON parser in particular, but I've already seen a fair share of MRs that contained many thousands of lines that no human would have ever written without stopping to ask "do I really need to do it?". LLMs aren't bothered.