With LLM-generated prose now ubiquitous, people are increasingly keen to identify its origins. While initial giveaways like em-dashes and the word “delve” have largely disappeared, researchers assert that AI models still exhibit numerous characteristic patterns when crafting text.
A recent study by the marketing firm Graphite investigated the writing tendencies of advanced AI models, pinpointing each model’s favored words and expressions. Although previous tells, such as em-dash usage, have been eliminated, models continue to rely on contrast-heavy sentence structures, with every new model version displaying its own distinct idiosyncrasies. The most surprising finding is the vast scope of these linguistic identifiers. Graphite uncovered 13,000 phrases that appeared at least twice as often in AI content compared to human writing — their established definition of an “AI tell.”
“It turns out that Claude models are actually getting closer to the human word distribution over time,” Greg Druck, Graphite’s chief AI officer, shared with TechCrunch. “And for the GPT models, it’s getting further away.”
Analyzing AI-generated writing at scale demanded a precise study design. Graphite commenced by gathering a corpus of 10,000 articles published before ChatGPT’s debut, which served as the human-authored control group. Subsequently, researchers instructed various AI models to rewrite these articles from summaries, aiming to minimize source bias. By comparing matched samples from both human writers and each model, they could assess the frequency of particular words and phrases in AI output, along with broader sentence construction patterns.
According to Graphite’s findings, Claude Opus 5.5’s most prominent tell is the adjective “dependable,” which appears 23 times more frequently than in human samples. While Opus 5.5 now avoids the “it’s not X, it’s Y” sentence structure, it still commonly employs the construction “is more than an X, it’s a Y.”
Crucially, Opus frequently emphasizes importance, using the phrase “this matters” 116 times more often than human writing, while “why X matters” occurs 92 times more often.
OpenAI’s Astra, conversely, presents a different set of identifiers. This model often refers to “another dimension” of its subject and tends to qualify assertions by stating an action “may provide” or “can provide” a specific advantage. Its most significant tell is what Graphite terms “corrective framing,” where a topic is defined as “not simply X” or offered as an alternative, “rather than relying on X.” Graphite’s research indicated these constructions were over 100 times more common in Astra-generated prose than in human writing.
Remarkably, all the leading AI labs seem to have addressed the models’ propensity for em-dash overuse. In Graphite’s collected samples, Opus 5.5 utilized the punctuation mark 99% less frequently than Opus 5. Astra now employs it 88% less often than human samples, and Gemini 3.1 Pro has nearly eliminated the em-dash from its generated text.
However, despite the individual tells evolving, Graphite notes that their overall prevalence remains largely consistent. “It’s not like the tells are decreasing,” Druck told TechCrunch. “They are managing to remove the most well-known tells, but other ones pop up. And every model version has its own.”
The persistence of these tells is surprising, given the labs’ emphasis on producing human-like writing styles. Upon the Opus 5.5 release, Anthropic proudly declared that the model “communicates more naturally than prior models,” with early users reportedly finding “its writing clearer and easier to follow.”
OpenAI made analogous assertions when introducing the GPT-6 versions of Sol and Luna, promising users could “expect to see more clarity, less jargon, [and] fewer odd turns of phrase.”
Nevertheless, Druck remains skeptical about the extent to which these labs can fully eradicate characteristic 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 explains. “These are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through.”