Large language models are statistics machines. They look at the previous word and predict the next word based on frequency and context.

Asking a large language model to write code means it will look across its vast library of code and write code similar to that. And when I tell you that most code on the internet is absolute garbage, I am being generous. On stackexchange, you will find one right answer and 100 wrong answers. Unless that LLM was first taught how to understand the ranking system of stack, it will think that the wrong answers are 100 times more valid.