I think that comes from confusing the human-inferred interiority of a fictional character versus the real-world nameless LLM author algorithm.

Suppose I make a black box program that generates a story about Santa Claus, a fictional character with lines about "love and kindness to all the children of the world" and claims to own a magical sleigh parked at the North Pole.

Does that mean I've created a program that has internalized and experiences love and kindness? Does my program necessarily have any geographic sense whatsoever about where the North Pole is?