A new study from marketing firm Graphite shows that AI models still have clear writing habits. The research compared prose from leading models with writing by people.
Old giveaways, such as heavy use of the em dash and the word "delve," have mostly faded. However, the study found many new patterns that can reveal machine-written text.
Graphite identified 13,000 phrases that showed up at least twice as often in AI content as in human content. The firm counts any phrase that meets that bar as a "tell."
How Graphite Ran the Study
Researchers started with 10,000 articles published before ChatGPT was released. These served as the human control group.
Different AI models then rewrote the articles using only summaries. The goal was to reduce bias from the source material.
With matching samples from people and each model, the team compared how often words and phrases appeared. They also looked at broader sentence patterns.
"It turns out that Claude models are actually getting closer to the human word distribution over time," Graphite chief AI officer Greg Druck told TechCrunch. "And for the GPT models, it's getting further away."
The Habits of Each Model
Anthropic's Claude Opus 5.5 uses the word "dependable" 23 times more often than human writers. That was its single biggest tell.
The model has dropped the "it's not X, it's Y" sentence form. It still tends to say something "is more than an X, it's a Y."
Opus 5.5 also likes to explain why things are important. It uses the phrase "this matters" 116 times more often than people, and "why X matters" 92 times more often.
OpenAI's Astra model has its own patterns. It often describes "another dimension" of a topic and hedges by saying something "may provide" or "can provide" a benefit.
Astra's biggest tell is what Graphite calls "corrective framing." Phrases like "not simply X" or "rather than relying on X" appeared more than 100 times as often as in human writing.
All the major labs appear to have cut back on em dashes. Opus 5.5 used them 99% less often than Opus 5, and Astra now uses them 88% less than human writers. Google's Gemini 3.1 Pro has almost completely removed the mark.
Druck said the total number of tells is mostly holding steady. "They are managing to remove the most well-known tells, but other ones pop up," he said. "And every model version has its own."
Anthropic said at launch that Opus 5.5 "communicates more naturally than prior models." The company said early users "found its writing clearer and easier to follow."
OpenAI made similar claims about the GPT-6 versions of Sol and Luna. It said users could "expect to see more clarity, less jargon, [and] fewer odd turns of phrase."
Druck doubts the labs can fully remove these patterns. "These are giant models with billions of parameters," he said. "They have some finite number of tests they can run, and things slip through."
