AI writing tells are the subtle phrases and sentence patterns that reveal when a piece of text was generated by a large language model. Even as models grow more sophisticated, researchers keep finding fresh giveaways. A new study from the marketing firm Graphite examined how frontier AI models write, and the results show that every model has its own recognizable habits.
The study’s scope was surprisingly broad. Graphite identified 13,000 phrases that appeared at least twice as often in AI-generated content as in human writing, which the researchers used as their definition of a tell. That number suggests detection is not about catching one obvious word. Instead, it is about noticing a cluster of small, repeated tendencies that add up to a distinct voice.
How the Study Measured AI Writing Tells
Studying AI writing at scale required a careful design. Graphite began with a corpus of 10,000 articles published before the release of ChatGPT, using them as a human-generated control group. Researchers then had different AI models rewrite those articles from summaries, aiming to remove as much source bias as possible.
With matching samples from humans and each model, the team could compare how often certain words and phrases appeared, as well as broader patterns in sentence construction. This side-by-side method is what makes the findings useful. Rather than guessing which phrases feel robotic, the study measured which ones actually appear far more often in machine output.
Opus 5.5’s Favorite Phrases
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. The model has moved away from the “it’s not X, it’s Y” construction, but it still leans on a similar pattern: saying something “is more than an X, it’s a Y.”
Above all, Opus 5.5 loves to explain why things matter. The phrase “this matters” shows up 116 times more often than in human writing, while “why X matters” occurs 92 times more often. These constructions give the prose a sense of importance, but they can also feel formulaic when they appear in nearly every paragraph.
Contrast and Correction as Habits
Contrast-heavy phrasing remains one of the most persistent AI writing tells. Models often define a topic by what it is not, then pivot to what it is. This pattern can be effective in small doses, but when it repeats, it starts to read like a template rather than a thought.
Astra’s Different Set of Tells
OpenAI’s Astra has its own distinct tip-offs. This model loves to describe “another dimension” of whatever it is discussing. It also tends to hedge claims by saying an action “may provide” or “can provide” a particular benefit.
Its biggest tell is what Graphite calls corrective framing, where a topic is defined as “not simply X” or offered as an alternative, “rather than relying on X.” According to the research, those constructions were more than 100 times more common in Astra-generated prose than in human writing.
Why the Em-Dash Tell Disappeared
Notably, all the frontier labs appear to have responded to the idea that models overuse em-dashes. In Graphite’s samples, Opus 5.5 used the punctuation mark 99 percent less often than Opus 5. Astra now uses it 88 percent less than human samples, while Gemini 3.1 Pro has almost completely eliminated the em-dash from its writing.
That shift shows labs can fix the tells they know about. Once a habit becomes widely mocked online, it becomes a target for retraining. The problem is that removing one tell does not remove the underlying tendency to fall into repetitive patterns.
The Overall Number of Tells Is Holding Steady
While individual tells change, Graphite says the overall number is mostly holding steady. Greg Druck, Graphite’s chief AI officer, explained that the tells are not decreasing. Models manage to remove the most well-known ones, but others pop up to replace them, and every model version has its own.
Druck also noted an interesting split. Claude models are actually getting closer to the human word distribution over time, while the GPT models are getting further away. That divergence suggests different labs are making different trade-offs in how they tune their writing styles.
Why Labs Struggle to Eliminate Tells
It is surprising that tells remain so persistent, given how much focus the labs place on human-like writing. In the Opus 5.5 release, Anthropic claimed the model communicates more naturally than prior models and said early users found its writing clearer and easier to follow.
OpenAI made similar claims when releasing the GPT-6 versions of Sol and Luna, saying users could expect more clarity, less jargon, and fewer odd turns of phrase. Yet the measured data shows that distinctive patterns survive these improvements.
Druck is skeptical about how much the labs can do to fully eliminate telltale constructions. He suggests the labs are less able to control some of these things than people might expect. These are giant models with billions of parameters, and with a finite number of tests they can run, things slip through.
What These AI Writing Tells Mean for Readers
For readers, the lesson is that detection is a moving target. The obvious signals fade as labs retrain their models, but new patterns take their place. Looking for a single word is less reliable than noticing a cluster of habits, such as constant importance framing, corrective contrast, and hedged benefits.
For writers and editors, the findings are a reminder to vary sentence structure and question whether every claim truly needs a line explaining why it matters. The most human writing is not the writing that avoids every flagged phrase. It is the writing that uses emphasis because the point deserves it, not because a pattern called for it.
