Not always, I’ll screen shot some if I see I get a new one, but I’ve definitely had some nonsense answers recently.

Also, it clearly isn't static. I played around with Gemini and when I got some clearly BS response, just repeating the prompt would often lead to a more correct result.

No, this is still a common occurrence and pretty well documented. LLMs give a random answer based on a dataset and programmed weighting towards certain types of responses. This is why you can give the same prompt to the same LLM repeatedly and get different responses each time, or different responses by slightly modifying the prompt, even if both prompts say essentially the exact same thing. There is no comparison or "learning" happening from user input. It doesn't think, rationalize, or memorize. This is just what LLMs are and how they work under the hood.

Anthropomorphizing LLMs is a bad idea, and trusting the output without manual verification is foolish. The LLM does not know or care about misinformation, it is just a software that analyzes a dataset and outputs that information with programmed noise for variance, and sometimes extra user ass-kissing added for flair.