Now that LLM-generated prose is everywhere, humanity is desperate for a way to sniff it out. Although early commands like em dashes and “dig in” are long gone, there are still many obvious habits that AI models rely on when writing prose, researchers say.
A new study released by marketing firm Graphite examined the writing habits of Frontier models and asked each model’s favorite words and phrases. Although older tells such as the use of em-dash have been discontinued, the model still relies on contrast-oriented construction, and each version of the model has its own quirks. The biggest surprise was that the tell’s range turned out to be very wide. Graphite found 13,000 phrases, or AI definitions of “tell,” that are at least twice as common in AI content than in human content.
“We found that the Claude model actually approached human word distribution over time,” Greg Druck, chief AI officer at Graphite, told TechCrunch. “And in the GPT model, it goes even further away.”
Studying AI-generated texts at scale required careful research design. Graphite started with a corpus of 10,000 articles published before the release of ChatGPT and served as a human-generated control group. The researchers then had different AI models rewrite the article from the summary, hoping to eliminate as much source bias as possible. By matching samples from both humans and each model, we were able to compare how often certain words and phrases appear in the AI’s writing, as well as broader patterns in sentence construction.
According to Graphite’s results, Claude Opus 5.5’s most distinctive feature is the word “reliable,” which occurs 23 times more frequently than in the human sample. Opus 5.5 avoids the sentence structure “It’s not X, it’s Y”, but still tends to say “It’s more than X, it’s Y”.
Above all, opus love to tell people why things are important, with the phrase “this is important” being used 116 times more often than humans write, and “why X is important” being used 92 times more often.
OpenAI’s Astra provides a different set of information. This model loves to explain the “other dimension” of whatever it’s talking about, and tends to avoid claims by saying that an action “might” or “could” provide a certain benefit. Its biggest feature is what Graphite calls “corrective framing,” where a topic is defined as “more than just X” or offered as an alternative “rather than relying on X.” According to Graphite’s research, these structures were more than 100 times more common in astra-generated prose than in human writing.
Notably, all of Frontier Labs seems to be reacting to the idea that the model makes heavy use of em dashes. In the Graphite sample, Opus 5.5 uses 99% less punctuation than Opus 5. Astra used 88% less punctuation than the human sample, while Gemini 3.1 Pro almost completely removed em-dashes from sentences.
But while individuals talk about change, Graphite said the overall numbers have remained largely stable. “Tellurium is not decreasing,” Druck told TechCrunch. “They’ve managed to remove the most well-known information, but other information shows up as well. And each model version is unique.”
It’s surprising that Tell is so persistent, given the lab’s emphasis on human-like writing. In its Opus 5.5 release, Anthropic boasted that the model “communicates more naturally than previous models,” and said that early users “found the writing to be clearer and easier to understand.”
OpenAI made a similar claim when it released GPT-6 versions of Sol and Luna, saying users can “expect more clarity, less jargon, (and) less weird phrasing.”
But Druck is skeptical about how far labs can go to completely eliminate telltale structures and phrases.
“The general hypothesis I have is that labs have less control over some of these things than you might expect,” Druck says. “These are huge models with billions of parameters. You can only run so many tests, and things will slip through the cracks.”
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
