As large language model-generated prose becomes ubiquitous, researchers are identifying persistent linguistic patterns that distinguish AI writing from human output. A new study by marketing firm Graphite reveals that while early markers like em-dash overuse have been largely addressed, frontier models continue to rely on specific phrases and syntactic structures at rates far exceeding human norms.
What Happened
Graphite analyzed writing habits across major frontier models by comparing a control group of 10,000 human-written articles published before ChatGPT’s release with AI-generated rewrites of those same summaries. The study identified 13,000 phrases that appeared at least twice as often in AI content as in human content. According to Greg Druck, Graphite’s chief AI officer, the research found that Claude models are generally moving closer to human word distributions over time, whereas GPT models are drifting further away.
Specific models exhibited distinct quirks. Claude Opus 5.5’s most frequent tell was the word “dependable,” appearing 23 times more often than in human samples. The model also heavily favored the phrase “this matters,” which occurred 116 times more frequently than in human writing, and “why X matters,” which appeared 92 times more often. While Opus 5.5 avoided the “it’s not X, it’s Y” construction, it frequently used the variant “is more than an X, it’s a Y.”
OpenAI’s Astra model showed different patterns, particularly a preference for “corrective framing” such as defining a topic as “not simply X” or offering alternatives with “rather than relying on X.” These constructions were more than 100 times more common in Astra-generated prose than in human writing. Astra also frequently used hedging language like “may provide” or “can provide” and often referenced “another dimension” of a subject. In contrast, em-dash usage has dropped significantly across the board, with Gemini 3.1 Pro nearly eliminating the punctuation mark from its output.
Why It Matters
The findings highlight the difficulty AI labs face in fully replicating human-like writing styles despite their efforts. Although Anthropic claimed in the Opus 5.5 release that the model “communicates more naturally” and that users found its writing “clearer,” and OpenAI made similar assertions for its GPT-6 versions, the study suggests that deep-seated statistical tendencies remain. Druck expressed skepticism about the labs’ ability to completely eliminate these tells, noting that with billions of parameters and finite testing capabilities, certain linguistic quirks inevitably “slip through.” This persistence of AI-specific phrasing offers a potential tool for developers and platforms seeking to identify machine-generated content in an increasingly saturated landscape.
The Bottom Line
While AI models have successfully reduced well-known tells like em-dash overuse, they have replaced them with new, model-specific linguistic habits. Graphite’s research indicates that the total number of detectable AI tells remains steady, with each new model version introducing its own unique set of overrepresented phrases and structures.