Tells found in Claude Opus 5.5: 2,548 (Source: Graphite, AI Tells study update, reported by VentureBeat (September 30, 2026))
A rule of thumb with limits
Many readers now treat an em dash as a sign of machine writing. It is a quick, free test. Graphite's new research shows why a test built on one habit is fragile.
Graphite, a digital marketing agency, found that Anthropic's new Claude Opus 5.5 almost never uses the em dash. Even so, the researchers counted 2,548 expressions in the model's articles that show up at double the human frequency or more. They include single words, short phrases and sentence templates.
The lesson here is simple. AI writing habits do not disappear. They move. A team that screens for one habit is testing for a pattern that differs by model and version.
What Graphite measured
The original study began with 10,000 human-written web articles. All were published before ChatGPT's public launch on November 30, 2022. Nine AI models then each wrote a matching piece for every one of those topics, which produced 90,000 machine-written articles to compare.
The method works in two steps. GPT-4.1 first summarised each human article. Each model then wrote a new article from that summary. This keeps the topics matched without copying the original.
A "tell" is a word, a short phrase or a sentence pattern that appears at least twice as often in AI text as in human text. Graphite also set minimum counts, so a tell had to show up in many separate articles.
Some tells are templates with a gap in the middle. Take "less like a _ and more like". It appeared 125 times in Claude Opus 5's articles and once in the human ones.
The em dash fell. The habits did not.
In Opus 5.5's articles, the em dash nearly vanished. Graphite counted 0.015 per 1,000 words, against 2.92 for Opus 5. That is a fall of about 99%.
In the original study, Opus 5 used em dashes at about the human rate. GPT-6 Astra used them at about one-eighth of the human rate. Gemini 3.1 Pro had nearly stopped. Graphite suggests the GPT and Gemini lines may have swung too far while trying to shed this well-known habit.
Other habits grew. Opus 5.5 uses "can help you" eight times as often as Opus 5 does, and "is especially helpful" 12 times as often. A model can drop one visible habit and pick up others.
Closer to human, but only in one sense
Graphite measured word-distribution divergence. It shows how differently a model and human writers use words overall. A lower score means closer to human.
Opus 5.5 moved toward the human sample, from 0.064 to 0.052. That is a 19% reduction. OpenAI's GPT-6 Astra moved the other way. It scored 0.109, against 0.101 for GPT-5.6 Sol.
Graphite also scored "mannered prose". Anthropic uses that term for writing that favours metaphor or flourish over direct statement. Opus 5.5 scored 10.57, down 37% from 16.75 for Opus 5. The human sample scored 6.65, and Astra scored 7.91. A lower score means plainer writing.
Graphite's CEO Ethan Smith told VentureBeat the pattern is not a steady climb. "It's just sort of different, and each one is different," he said.
Graphite's chief AI officer, Gregory Druck, addressed quality directly. "I don't think we're necessarily passing judgment on that here in this work," he said.
VentureBeat's reading of the metric is narrower still. Opus 5.5's overall word use comes closer to the human articles Graphite chose. That does not show readers would mistake an article for human work. It also does not show the content is more accurate, original or useful.
Each model family has its own habits
Across the nine models in the base study, Graphite found 12,877 unique tells. Of these, 65% belong to a single model family.
Updates change habits too. Take the two newest releases in a family. Between 55% and 72% of the tells show up in only one of them.
This matters for any policy that lists "AI phrases" to avoid. The list would need to differ by vendor. It would also need frequent updates, because versions change.
Limits to keep in mind
Graphite used one fixed prompt and web articles only. Different prompts or settings, such as chat, could produce different tells. The mannered-prose scores described above were assigned by Claude Opus 5 on a subsample of 1,000 topics, so some evaluator bias is possible.
Graphite also does not recommend judging authorship from individual tells. For that task it recommends an AI detector.
Questions for your content and risk teams
First, does any policy rely on spotting phrases or punctuation to catch AI text? If so, ask what it does when those habits change.
Second, how does the team check content for accuracy and originality? Style cannot answer those questions.
Third, which models do your vendors and agencies use, and which versions? Habits differ by family and by release.
A missing em dash tells you nothing about who wrote a page. Review the substance, not the punctuation.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Graphite.
