Study maps the telltale phrases each AI model can't shake

Now that LLM-generated prose is everywhere, people are eager to spot it. Researchers say old tells like em-dashes and the word “delve” are mostly gone, but frontier models still fall back on their own habits. A new study from marketing firm Graphite mapped each model’s favorite words and phrases, built from a corpus of 10,000 articles published before ChatGPT’s release as a human-written control group; different AI models then rewrote those same articles from summaries to strip out source bias. Comparing matching samples, Graphite found 13,000 phrases that were at least twice as common in AI content as in human content — its definition of a “tell” — with each model carrying its own contrast-heavy quirks.

According to Graphite’s results, Claude Opus 5.5’s biggest tell is the word “dependable,” which appears 23 times more often than in human samples. While Opus 5.5 has dropped the “it’s not X, it’s Y” construction, it still tends to say something “is more than an X, it’s a Y.” Above all, Opus favors telling readers why things matter, using the phrase “this matters” 116 times more often than human writing, and “why X matters” 92 times more often.

OpenAI’s Astra has a different set of tip-offs: it likes to describe “another dimension” of a topic and hedges claims by saying an action “may provide” or “can provide” a given benefit. Its single biggest tell is what Graphite calls “corrective framing,” where a topic is cast as “not simply X” or offered as an alternative “rather than relying on X” — constructions Graphite found more than 100 times more common in Astra’s writing than in human text.

All of the frontier labs appear to have responded to criticism that models overuse em-dashes: Opus 5.5 used the mark 99% less often than Opus 5, Astra cut its use 88% versus human samples, and Gemini 3.1 Pro has nearly eliminated it from its writing. But Graphite’s chief AI officer, Greg Druck, says the overall number of tells is holding roughly steady even as individual quirks shift, with labs scrubbing the best-known ones only for new ones to surface in each model version. “A general hypothesis I have is that the labs are less able to control some of these things than you might expect,” Druck told TechCrunch.

Druck also said Claude models are drifting closer to the human word distribution over time, while GPT models are drifting further away. That persistence is notable given how much the labs tout natural-sounding writing: Anthropic has said Opus 5.5 communicates more naturally than prior models, with early users finding its writing clearer and easier to follow, while OpenAI has made similar claims for its GPT-6 models.

Why it matters

Graphite’s findings suggest that as AI writing sheds its most obvious quirks, spotting it will increasingly require tracking narrower, model-specific patterns rather than a handful of universal giveaways — a moving target for anyone trying to tell human and AI-generated text apart.