Well, it depends on what you're asking the LLM to generate. If I ask an LLM to write me an essay in the style of a caveman, the LLM simulates a redditor simula

Well, it depends on what you're asking the LLM to generate. If I ask an LLM to write me an essay in the style of a caveman, the LLM simulates a redditor simulating a fictional caveman (who speaks english but with short words and bad grammar). After all, that's what was in the training data. If I ask the same LLM to write me an essay in the style of a Harvard professor - who's to say the same thing doesn't happen? So I can see how, when training a model, you might want it to know what's really written by a Harvard professor, and what merely professes to be.