The most common LLM privacy advice — scrub personal data out of the prompt before the model sees it — aims at the wrong risk and quietly makes the model dumber. Here's what the research actually shows, and where the masking really belongs.