It's not an add-on feature. The LLM produces something with the best score it can. Things that increase the score:

• Things appropriate to the tokens in the request
• Things which look like what it's been trained on.

So that includes:

• Relevant facts
• grammatically correct language
• friendly style of writing
• etc

If it has no relevant facts then it will maximise the others to get a good score. Hence you get confidently wrong statements because sounding like it knows what it's talking about scores higher than actually giving correct information.

This process is inherent to machine learning at its current level though. It's like a "fake it until you make it" person, who will never admit they're wrong.