Graphite Study: Opus 5.5 Is Distinguished From Humans by 2,548 Language Patterns
Graphite's updated AI Tells study, shared with VentureBeat before publication in early October 2026, shows that the most notorious AI language marker — the em dash — has been nearly eliminated from the newest flagship models. But it has been replaced by thousands of new ones: In Claude Opus 5.5 alone, 2,548 words, phrases, and sentence patterns are mapped that are at least twice as common as in human-written articles. At the same time, the numbers show that the models are moving in opposite directions — Anthropic closer to human vocabulary, OpenAI further away.
What's New
The study from the marketing firm Graphite is an updated version of their AI Tells analysis, shared with VentureBeat ahead of planned publication on Thursday in early October 2026 [1]. The update adds the newest flagship models — Anthropic's Claude Opus 5.5 and OpenAI's GPT-6 Astra, which VentureBeat reports were launched earlier in the month — to a comparison across ten models.
The core of the analysis: Graphite defines a "tell" as a word, phrase, or sentence pattern that is at least twice as common in AI-generated text as in comparable human-written text. In total, the team identified around 13,000 such phrases across the models [3]. For Opus 5.5 alone, 2,548 patterns apply [1]. The figures cover different frameworks — 13,000 is the total across ten models, 2,548 is the per-model figure for Opus 5.5 — and the sources do not specify exactly how they relate.
The Method
Graphite started with a corpus of 10,000 articles published before ChatGPT launched as the human control group. They then had various AI models rewrite the articles from summaries, to eliminate as much source-side bias as possible — that is, differences in source coverage should not affect the comparison [2]. The comparison spans 9,974 comparable topics across ten models [1].
The Old Tells Are Disappearing
The most striking change is the em dash. Graphite counts 0.015 em dashes per 1,000 words in Opus 5.5 texts, versus 2.92 in Opus 5 — a drop of roughly 99 percent [1]. OpenAI's GPT-6 Astra uses the em dash 88 percent less than human samples, while Google's Gemini 3.1 Pro has removed it almost entirely [3].
Graphite's so-called "mannered prose" score, an aggregate measure of affected writing style, fell 37 percent for Opus 5.5 — from 16.75 to 10.57. It still sits above the human figure of 6.65 [1].
The New Tells
But the elimination of old markers has created new ones. For Opus 5.5, the biggest tells are about telling the reader why things matter: The phrase "this matters" occurs 116 times more often than in human writing, and "why X matters" 92 times more often [2]. Opus 5.5's largest single tell is the word "dependable" — TechCrunch and Newsbytes report it is 23 times more common than in human samples, while Gizmodo reports 26 times [2][3][4]. The discrepancy is not resolved in the sources.
Astra has a different signature pattern. Graphite calls it "corrective framing": A topic is defined as "not simply X" or presented as an alternative, "rather than relying on X." According to the research, such constructions are over 100 times more common in Astra-generated text than in human writing [2]. Astra also hedges conspicuously often with "may provide" and "can provide" [2]. And according to Gizmodo, OpenAI's flagship uses the phrase "does not establish" at an absurd frequency — 275 times more common in Astra texts than in Claude texts [4].
The Models Are Moving in Different Directions
Despite old tells disappearing, the models are not collectively moving toward human language. Graphite measures the distance between each model's vocabulary distribution and the human sample: Opus 5.5 scores 0.052, down from 0.064 for Opus 5 — a 19 percent reduction. GPT-6 Astra moves the opposite way: 0.109, up from 0.101 for GPT-5.6 Sol [1].
Note that the two measurements concern similar protocols, not the same scale with a guaranteed direction — it is not the case that one model is objectively "better" than the other on all dimensions. But the direction is clear: The Anthropic model is approaching human vocabulary, the OpenAI model is moving away.
The Human-Likeness Claim Meets the Numbers
The contrast with the labs' own claims is pointed. In the launch of Opus 5.5, Anthropic claimed that the model "communicates more naturally than previous models" [2]. OpenAI has made similar claims about human-like writing for GPT-6 Sol and Luna [2].
Greg Druck, Graphite's chief AI officer, is skeptical that the labs can fully eliminate the tells: "A general hypothesis I have is that the labs are less able to control some of these things than you might expect," he says. "These are giant models with billions of parameters. They have a finite number of tests they can run, and things slip through" [2].
Why It Persists
Druck's explanation points to a structural weakness: Even though the labs explicitly try to eliminate known tells — the em dash decline shows that it actually happens — testing capacity is not proportional to the models' complexity. When the em dash is eradicated, new patterns emerge in the same way: "this matters," "corrective framing," "does not establish." The total number of tells is largely unchanged across model generations [3] — the form shifts, the volume does not.
Caveats
Graphite is a marketing firm, something Gizmodo points out explicitly [4]. The firm's business model has a clear commercial interest in AI detectability, which gives reason for source-critical restraint.
In addition, the study does not measure whether readers actually mistake AI articles for human-written ones, and it does not say whether closer stylistic adjustment means better or more accurate content — VentureBeat notes this limitation via Druck [1]. It should also be emphasized that all the figures here come from press coverage of the study, not from Graphite's own published report.
Sources
[1] VentureBeat – "Claude Opus 5.5 uses em-dashes 99% less often and sounds more human — but still exhibits 2,548 AI writing tells" – https://venturebeat.com/data/claude-opus-5-5-uses-em-dashes-99-less-often-and-sounds-more-human-but-still-exhibits-2-548-ai-writing-tells
[2] TechCrunch – "Opus 5.5 loves to tell you 'this matters' (and other AI writing tells)" – https://techcrunch.com/2026/10/01/opus-5-5-loves-to-tell-you-this-matters-and-other-ai-writing-tells/
[3] Newsbytes – "Study finds 13,000 phrases that could reveal AI-generated writing" – https://www.newsbytesapp.com/news/science/study-reveals-ai-tells-in-writing-like-overusing-certain-words/story
[4] Gizmodo – "'This Matters': Researchers Identify Thousands of New Tells in AI Writing" – https://gizmodo.com/this-matters-researchers-identify-thousands-of-new-tells-in-ai-writing-2000820447