I think its because LLMs are often trained with data from code tutorial sites and forums like stackoverflow, and not always production code