Is it more likely to be clear and reliable if it is AI-written, or are features associated (both directly and by correlation) with clear writing increasingly misperceived as “AI tells” because they are also favored in LLM training?