AI-generated prose has become so widespread that people are increasingly looking for ways to identify it. Although early giveaways like em-dashes and the word "delve" have largely disappeared, researchers say AI models still have plenty of recognizable writing habits they can't seem to shake. A recent study by the marketing firm Graphite examined how frontier models write, identifying each model's preferred words and phrases. While older tells such as em-dash usage have been eliminated, models continue to rely on contrast-heavy constructions, and each version of a model displays its own distinct quirks. What's most striking is just how extensive the list of tells has become. Graphite identified 13,000 phrases that appeared at least twice as frequently in AI content compared to human content—their threshold for what counts as a "tell." "And for the GPT models, it's getting further away."

EMBED_PLACEHOLDER_0

Analyzing AI-generated writing at scale demanded a rigorous study design. Graphite began with a corpus of 10,000 articles published before ChatGPT's release, which served as the human-generated control group. Researchers then had various AI models rewrite those articles from summaries, aiming to minimize source bias as much as possible. With matched samples from both humans and each model, they could compare how frequently particular words and phrases showed up in AI writing, along with broader patterns in sentence structure. Based on Graphite's findings, Claude Opus 5.5's most notable tell is the word "dependable," which appears 23 times more often than in human samples. Although Opus 5.5 has moved away from the "it's not X, it's Y" sentence pattern, it still tends to say something "is more than an X, it's a Y."

More than anything, Opus loves explaining why things matter, using the phrase "this matters" 116 times more frequently than human writing, while "why X matters" shows up 92 times more often. OpenAI's Astra has a different set of giveaways. This model enjoys describing "another dimension" of whatever subject it's addressing, and tends to soften claims by saying an action "may provide" or "can provide" a particular benefit. Its most prominent tell is what Graphite calls "corrective framing," where a topic gets defined as "not simply X" or presented as an alternative, "rather than relying on X." According to Graphite's research, those constructions appeared more than 100 times more frequently in Astra-generated prose than in human writing. Notably, all the frontier labs appear to have responded to the criticism that models overuse em-dashes. In Graphite's samples, Opus 5.5 used the punctuation mark 99% less often than Opus 5. Astra now uses it 88 percent less than human samples, while Gemini 3.1 Pro has nearly eliminated the em-dash from its writing entirely. But even as individual tells shift, Graphite says the overall count remains largely stable. "They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own."

EMBED_PLACEHOLDER_1

It's surprising that tells persist so stubbornly, given how much the labs focus on achieving human-like writing styles. In the Opus 5.5 release, Anthropic bragged that the model "communicates more naturally than prior models," claiming early users "found its writing clearer and easier to follow."

OpenAI made similar assertions when releasing the GPT-6 versions of Sol and Luna, saying users could "expect to see more clarity, less jargon, [and] fewer odd turns of phrase."

But Druck is doubtful about how much the labs can actually do to fully eliminate telltale constructions or phrases. "A general hypothesis I have is that the labs are less able to control some of these things than you might expect," Druck says. "These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through."