A couple weeks ago, I was curious to ask an LLM how it would distinguish fictional works like Harry Potter from factual non-fiction, like news reports or research. So I asked one or two of them how they knew.

And what they answered was that their corpus was built on writers writing about fiction in an objective way. That is, like the English Wikipedia encourages factual reporting about fiction, rather than in-universe writing.

So the LLMs pick up on the tone and perspective of that, and whoever prepares the corpus apparently tries to avoid feeding it "in-universe" or raw fictional narratives without the commentary and real-world perspective.

In other words, LLMs have no idea what is fact or fiction. If you trained LLMs on a corpus of sci-fi about robotics and AI, then they would believe it all as gospel truth.

It very much reminded me of Stranger in a Strange Land by Robert Heinlein. Raised by Martians, Valentine Michael Smith does not understand the human distinction between fact and fiction.