NPR: Pangram AI Detector Flagged 37 of 40 Mixed Human-AI Papers in Peer-Reviewed Study, But Human Writers Still Falsely Accused

Windsor Johnston reports that Pangram, cited in a June International Journal for Educational Integrity study as the most reliable detector, correctly flagged 37 of 40 mixed human-AI papers and 37 of 40 human-rewritten AI papers — yet its own CEO concedes misses on short samples, and competing tools still contradict each other on identical text. Writers quoted say they are being falsely accused when they use em dashes, 'delve', rule-of-three lists or adverbial openings like 'notably', pushing some to deliberately degrade their prose to avoid classifiers.