LLM writing is like AI-generated photos in that you don't notice the good instances of LLM writing, i.e. you don't know your false negative rate.