Vendor claims of near-perfect accuracy did not survive contact with actual student writing. Independent testing of commercial AI detectors consistently shows performance between 60 and 80 percent once text is paraphrased or lightly edited — and OpenAI's own tool correctly identified AI-generated content just 26% of the time before the company shut it down. The gap between advertised accuracy and operational accuracy is not a calibration problem. It is a category problem: these tools measure statistical proximity to a distribution, not authorship.