A new study by the marketing firm Graphite analyzed which words and constructions regularly reveal frontier large language models’ (LLMs) prose. The researchers found that although early, obvious tells (like frequent em‑dash use) have been reduced, models still revert to characteristic phrasing and sentence patterns — and each model version shows its own identifiable habits.
Methodology
Graphite began with a corpus of 10,000 human‑written articles published before the release of ChatGPT as the control group. To limit source bias, the team created summaries of those articles and asked different AI models to rewrite them from those summaries. With matched samples from humans and each model, Graphite compared the frequency of particular words and phrases and examined broader sentence construction patterns.
Key findings
-
Graphite identified 13,000 phrases that appeared at least twice as often in AI‑generated text as in human text; the firm treats those as model “tells.”
-
Anthropic Opus 5.5: the single strongest signal was the word "dependable," which appeared 23 times more often than in human samples. While Opus 5.5 largely avoids the simple "it's not X, it's Y" frame, it still favors comparative constructions like "is more than an X, it’s a Y." Notably, Opus 5.5 disproportionately uses phrases that emphasize importance: "this matters" occurred 116 times more often, and constructions like "why X matters" were 92 times more common than in human writing.
-
OpenAI Astra: Astra showed a different profile. It frequently describes an "another dimension" of a topic and hedges benefits with verbs such as "may provide" or "can provide." Graphite highlighted a dominant "corrective framing" in Astra outputs — defining topics as "not simply X" or presenting alternatives "rather than relying on X" — constructions that were more than 100 times more common in Astra than in the human control samples.
-
Em‑dashes: in response to earlier critiques, frontier labs have substantially reduced em‑dash usage. In Graphite’s samples, Opus 5.5 used em‑dashes 99% less often than human samples, Astra used them 88% less often, and Google Gemini 3.1 Pro has nearly eliminated the em‑dash from its prose.
Interpretation
Greg Druck, Graphite’s chief AI officer, said the set of tells is not simply shrinking: developers have succeeded in removing the most well‑known markers, but new ones tend to emerge, and each model version develops its own profile. Druck also observed that Claude (Anthropic) models are moving closer to human word distributions over time, whereas GPT family models are, by his account, diverging.
Graphite’s findings are notable alongside the claims from model developers: Anthropic promoted Opus 5.5 as communicating "more naturally than prior models," with users finding its writing clearer and easier to follow. OpenAI made similar statements about the GPT‑6 based Sol and Luna releases, promising more clarity, less jargon, and fewer odd turns of phrase. Despite those goals, Druck expressed skepticism that labs can fully eliminate telltale constructions: these are enormous models with billions of parameters and only a finite number of tests, so some patterns slip through.
Implications
The study suggests that while obvious stylistic signals can be reduced, model‑specific linguistic habits persist and evolve. That is important for developers of AI detection tools, publishers, and readers seeking to identify machine‑generated prose — but it also implies detection strategies must continually adapt as models and versions introduce new tells.
