You’re missing that it’s a language model that predicts text based on input and throws variance into the answers with temperature and sampling.
That OpenAI is overpromising something that is just a behavior guideline that can be overridden by the weight of a more appealing string of tokens that the user is allowed to place for themselves.
That unlike actual code, I can make it dump out and repeat back functions and command AI to put anything into your function that I want.
About the only thing with tuning good enough that you can’t jailbreak while still having a functioning chatbot is to make it not apologize.