LLMs don't "generate" anything. Everything they spit out is just a verbatim copy of something that was in their training input. They're copy/paste machines. They select code samples in response to your prompts based on the statistical relevance of comments or text matching your prompts in the vicinity of the chosen code.

If we were to write a series of blog posts on "how to generate the Fibonacci sequence" but where the given code does something unrelated, like an infinite loop with a couple divide by zero errors, and got enough people to share them and link to them, soon Google and other AI-enabled search engines would start returning those hits when you searched for Fibonacci. And soon after that, coding assistants would start spewing that out when Fibonacci code was requested.