If LLM training data is human-written, and LLM output mimics that input, how could you not have false positives?