It's even worse. When working with someone else's code e.g. StackOverflow there's a reputation system gating people from the platform and incentivizing them to provide correct answers. You can reasonably expect that someone else's code has at least been thought through to some extent to solve the problem at hand, and very likely tested.

With LLM-generated code, especially ChatGPT-style decoder models, none of that is true. All of the posts and comments I see about it here seem to be anecdotes "it can do all of my job for me" yet asking it to write the simplest code creates several issues on my end.

Personally I think a model geared towards code generation isn't an unsolvable task; the Spider dataset was released some time ago (text to SQL task) and the winning approach there was no fanciness on the model side, but rather to just test all the output queries to ensure it's at least valid SQL. That got a 20%+ boost in accuracy.