Almost AI, Almost Human: The Challenge of Detecting AI-Polished Writing