Comment by sidpatil

Someone here mentioned a whole ago that the labs deliberately haven't tried to train these characteristics out of their models, because leaving them in makes it easier to identify, and therefore exclude, LLM-generated text from their training corpus.