Much ado has been made of “AI tells” in writing—so much so that an entire micro-economy of so-called AI detectors has cropped up and is somehow worth millions of dollars. A new study from Graphite (a marketing company, so you know, take it with a grain of salt) has identified about 13,000 words and phrases that it believes are likely indicators that an AI model produced text—and warns that in at least one case, the tells might be getting harder to spot.
Gone are the days when excessive em dashes were the primary flag of AI-generated content (and thank god, some of us need those). Now, according to Graphite, the biggest single tell is the phrase “this matters,” which appears a whopping 116 times more often in AI-generated content than in the human-produced samples used for comparison.
“This matters” is the giveaway that plagues Claude Opus 5.5, Anthropic’s flagship model, the most at the moment. It is also more likely to use the phrases “looking ahead the,” which appears in Claude outputs 40 times more frequently than it does in human material; “why [blank] matters,” which crops up 92 times more often; and the word “dependable,” which is used by Claude 26 times as much as human counterparts.
OpenAI’s Astra model has its own set of tells, per the study. For instance, OpenAI’s flagship model uses the phrase “does not establish” with a ridiculous frequency. The phrase is 275 times more common in Astra text than it is in Claude’s outputs.
In general, the study found that Opus 5.5 qualifies claims less often than Astra does, and is far more likely to use superlatives.
But perhaps the most noteworthy difference between Opus and Astra is the fact that Opus seems to be eliminating its tells. According to the study, Claude-produced text is getting closer to human writing over time. It even seems to be correcting (or possibly overcorrecting) for previous giveaways. Per the study, Opus 5.5 uses em dashes 99% less often than Opus 5 did.
Of course, it’s a bit of a game of whack-a-mole. Every time that an AI model addresses one of its apparent tells, other ones crop up. But the general trend, at least for Opus, seems to be moving toward less distinction between its outputs and generic human text. The same can’t be said about Astra, which is actually trending toward less human-like writing.
