Old AI writing tells are gone, but new verbal tics keep appearing
Vendors have stamped out famous tells like em-dashes, yet every model version grows its own high-frequency phrases, so no single feature reliably flags AI text.
ImportanceLocalEvidenceE2 unreplicatedWrite-upQuick
Readers can now know that the famous old AI writing tells have been stamped out, but after comparing word frequencies Graphite found 13,000 phrases at least twice as common in AI text, and every model version grows new verbal tics of its own.
People previously relied on single features like em-dashes to flag AI text, but vendors have now largely eliminated those old tells, so such single-feature detection no longer works.
Graphite designed and ran the measurement itself: it used 10,000 pre-ChatGPT articles as a human control, had models rewrite them from summaries, and compared word frequencies. Claude Opus 5.5's biggest tell is "this matters," 116 times more frequent than in human writing; OpenAI's Astra favors "not simply X" corrective framing, over 100 times more common.
Graphite chief AI officer Greg Druck told TechCrunch that models only remove the best-known tells, new ones keep popping up, and every version has its own.