Graphite researchers have identified 13,000 phrases that appear at least twice as often in AI-generated text as in human writing. These patterns remain visible despite the removal of early markers like em-dashes.
The study
The marketing firm began with a corpus of 10,000 articles published before ChatGPT arrived. These served as the human baseline. Researchers then asked various AI models to rewrite summaries of those articles. This method aimed to remove source bias while allowing a direct comparison of word frequency and sentence structure.
Claude Opus 5.5
Claude Opus 5.5 uses the word “dependable” 23 times more frequently than humans do. The model has largely stopped using the “it’s not X, it’s Y” construction. Instead, it prefers phrasing like “is more than an X, it’s a Y.”
The most obvious trait is a habit of justifying relevance. Opus 5.5 uses the phrase “this matters” 116 times more often than human writers. The variation “why X matters” appears 92 times more often.
OpenAI Astra
OpenAI’s Astra model focuses on describing “another dimension” of a subject. It also hedges claims by stating an action “may provide” or “can provide” a benefit.
The firm calls its most common pattern “corrective framing.” This involves defining a topic as “not simply X” or suggesting an alternative “rather than relying on X.” These constructions appear more than 100 times more often in Astra output than in human writing.
Punctuation habits
Every major lab has reduced the use of em-dashes after criticism that models overused them. Opus 5.5 uses the punctuation 99% less often than its predecessor, Opus 5. Astra is 88% below human usage levels. Gemini 3.1 Pro has almost completely removed the em-dash from its output.
Why the patterns persist
Graphite reports that the total number of tells remains steady. Greg Druck, chief AI officer at Graphite, told TechCrunch that models remove well-known markers only for new ones to appear. Every version carries its own quirks.
This persistence is surprising given the labs’ public goals. Anthropic stated that Opus 5.5 “communicates more naturally than prior models,” with early users finding the writing “clearer and easier to follow.”
OpenAI made similar claims for the GPT-6 versions of Sol and Luna. They said users could “expect to see more clarity, less jargon, [and] fewer odd turns of phrase.”
Druck remains doubtful that labs can fully eliminate these markers. “These are giant models with billions of parameters,” he says. “They have some finite number of tests they can run, and things slip through.”
What it means
Writers and editors should not rely on a single marker to detect AI text. The specific phrases change with every update, but the volume of distinctive patterns stays constant. Checking for the presence or absence of em-dashes is no longer a reliable test.
