The phrases that give away AI writing, measured across models

Everyone has a feel for what AI writing sounds like, but a study covered by TechCrunch tried to measure it. The marketing firm Graphite had frontier models rewrite 10,000 articles that predate ChatGPT, working only from summaries, then compared the output to the human originals. It found roughly 13,000 phrases that show up at least twice as often in machine text as in human text.

The tells are specific and, at times, funny. Claude Opus 5.5 reaches for "dependable" about 23 times more often than people do, and leans on "this matters" 116 times more often, along with the construction "more than an X, it's a Y." OpenAI's Astra keeps saying "another dimension" and hedging with "may provide" or "can provide," and it uses corrective framing, defining something as "not simply X," more than 100 times above the human rate. One habit is fading on purpose: Opus 5.5 cut its em-dash use by 99% against the previous version, and Gemini 3.1 Pro has nearly dropped the character.

What makes the piece useful rather than a list of gotchas is Graphite's observation that the tells keep moving. As labs train out one set of quirks, a new set appears with the next model, which suggests these patterns are a side effect of scale that no one fully controls. You can read the full writeup at TechCrunch.

Why it matters

If you edit text that models help produce, this is a cheap checklist for the current crop of tells, and a reminder that any detector tuned to today's phrases will be stale by the next release, so do not trust one as proof either way.