Pangram's AI Detector Is Reshaping What Human Prose Looks Like

AI detectors are now shaping how humans write. Bloomberg's Morgan Meaker reports that authors and students are stripping em dashes, dropping the word 'delve', avoiding 'not X, but Y' phrasing, and seeding typos into their drafts so classifiers like Pangram do not flag their prose as machine-generated.

The habits track the data. Pangram's own tallies, drawn from millions of documents, show AI uses em dashes at 17 per 10,000 words against 2 in human text, 'delve' at 30 against 3, the rule-of-three pattern at 19 against 5, and antithetical constructions like 'not just X but Y' three times more often than people do. The em dash alone, Pangram says, has pushed labs to train models not to use it.

In a study published in June in the International Journal for Educational Integrity, researchers tested four detectors across 160 academic papers. Pangram was the most consistently accurate, correctly identifying AI use in 37 of the 40 papers that mixed human and machine text, and in 37 of the 40 where a human had rewritten AI-generated material. That accuracy is also the trap for honest writers. Ghostwriter Sally McGraw, who was falsely flagged, told NPR's companion piece, 'It's like trying to fool a machine into thinking I'm not a machine.'

The brittleness is the other half of the story. Writers who have tested Pangram have shown that swapping a single em dash for a colon in the same passage can flip a verdict from 100% AI to 100% human, which means an actual cheater can bypass with one keystroke what a careful human cannot stop triggering. Science blogger Janelle Shane, who ran her own writing through the tool, was flagged as AI with 98.7% confidence.