Much the same as 'arguments' I can have with LLM's about things where I'm the expert and I know it's wrong, but it will justify its position to the end because it's trained on common misconceptions that exist among less-expert people.

The idea that's been floating around in my head for the last few years is something like "it's being trained by the data produced by people, it's going to have many human flaws as a result"