Graphite finds 13,000 phrases that flag AI writing, including 'this matters'

Graphite's research team cataloged 13,000 phrases that appear at least twice as often in AI-generated prose as in human writing, according to a new study that maps the writing habits of frontier models including Claude Opus 5.5, OpenAI's Astra, and Gemini 3.1 Pro.

The marketing firm released the findings this week. It treats any phrase that crosses that 2x frequency threshold as a "tell," a fingerprint a model can't quite shake, and the count underscores just how broad the AI-writing signature has become across the industry.

What did the study actually do?

The team built a strict before-and-after design. Researchers started with a corpus of 10,000 articles published before ChatGPT's release, a human-generated control group grounded in writing from 2022 and earlier. They then had each frontier model rewrite those same articles from summaries rather than full text, a method intended to strip out the lexical fingerprint of the original sources and isolate what the model wanted to add on its own.

With matched human and AI samples in hand, Graphite compared word frequencies and broader sentence-construction patterns. The 13,000 phrases that crossed the 2x bar across any single model became the study's headline metric.

"It turns out that Claude models are actually getting closer to the human word distribution over time," Greg Druck, Graphite's chief AI officer, told TechCrunch. "And for the GPT models, it's getting further away."

Which phrases give Claude away?

Claude Opus 5.5 has a clear favorite among both its verbs and its meta-commentary. Its largest single tell is the word "dependable," which surfaces 23 times more often than in human samples. The model also reaches for the construction "is more than an X, it's a Y" far more than people do, a residual version of the once-common "it's not X, it's Y" template.

Above all, Opus loves to tell the reader why things matter. The phrase "this matters" appears 116 times more often in Opus writing than in human writing. Its close cousin "why X matters" appears 92 times more often. Together the two phrases form a meta-rhetorical tic, a model habitually announcing the significance of the very point it is making.

Which phrases give OpenAI's Astra away?

Astra's tells cluster around qualification and correction. Graphite groups them under "corrective framing": sentences that define a topic as "not simply X" or offer it as an alternative "rather than relying on X." Those constructions appeared more than 100 times as often in Astra-generated prose as in human writing.

Astra also hedges concrete claims. It favors the verbs "may provide" and "can provide" when describing what an action will do for a reader, and it leans on the phrase "another dimension" as a soft escalator for an idea. The pattern suits a model tuned to sound careful, and the resulting prose carries a recognizable diplomatic cadence.

Are the old em-dash tells really gone?

The signature overuse of em-dashes that marked early ChatGPT output has largely disappeared from frontier models. Graphite's samples show Opus 5.5 using em-dashes 99 percent less often than Opus 5. Astra now deploys them 88 percent less than human writers. Gemini 3.1 Pro has almost entirely removed the dash from its output.

That change reflects direct intervention. Em-dashes became the first viral AI-tell, and the labs clearly responded. But each time the most visible tell is engineered out, new ones surface elsewhere in the distribution.

"It's not like the tells are decreasing," Druck told TechCrunch. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own."

Why are these patterns so persistent?

The persistence runs against the explicit marketing claims of the labs releasing these models. Anthropic's Opus 5.5 release notes said the model "communicates more naturally than prior models" and reported that early users "found its writing clearer and easier to follow." OpenAI's GPT-6 variants Sol and Luna promised users they would "expect to see more clarity, less jargon, [and] fewer odd turns of phrase."

Druck is skeptical that the labs can keep scrubbing tells indefinitely. "A general hypothesis I have is that the labs are less able to control some of these things than you might expect," he said. "These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through."

That gap between vendor claims and statistical reality is what makes the study matter. Every frontier-model release is now accompanied by language promising a more human voice, and every release quietly introduces its own renewed signature.

What does this mean for AI detection?

The snapshot carries practical stakes for the growing AI-detection market. With 13,000 cataloged tells across current frontier models, detection vendors can no longer rely on a small list of red-flag words. The em-dash, "delve," and other one-word markers are out. Sentence-level patterns such as corrective framing, the "this matters" tic, and hedged benefits are in.

For newsrooms, the study sharpens the practical question of disclosure. Static word lists are no longer enough. Editors who want to flag undisclosed AI drafting now have to inspect construction, not just vocabulary, and the tells will differ by which model touched the file.

What happens next?

Graphite's next study cycle will follow the same playbook: take a fresh batch of pre-ChatGPT articles, run them through whatever frontier models are current at the time, and re-measure the tells. Each release from Anthropic, OpenAI, or Google becomes another data point in a slowly shifting distribution.

Until the labs can demonstrably close the distribution gap, and Druck, for one, doesn't expect that soon, the AI-writing signature will remain a moving target. The tells will move. They will not disappear.