The Cheating Detection Threshold

Your institution has licensed an AI-cheating-detection tool that returns a 'likelihood of AI authorship' score from 0 to 100. You have to set the threshold above which submissions are flagged for review. At 50%, roughly 18% of submissions are flagged, with documented false positives that fall hardest on English-language learners. At 80%, only about 3% are flagged, but you miss most actual cases. Faculty are demanding a number. The number is a value judgment dressed as a technical setting.