Graphite Finds 13,000 Phrases AI Uses More Often Than Humans
Graphite identified 13,000 phrases that language models use at least twice as often as humans. The list can help editors review AI-assisted writing, but it cannot establish who wrote a particular article.
According to a Graphite study reported by Habr on October 3, 2026, language models use 13,000 phrases at least twice as often as people do in the texts used for comparison. Each model had its own characteristic turns of phrase, which also changed between versions. The list can help when editing AI-assisted writing, but it is not a test of who wrote a particular article.
How Graphite Compared the Texts
The control group consisted of 10,000 articles published before ChatGPT was released. The researchers prepared summaries of those articles and asked different models to write new pieces based on them. This was intended to reduce the influence of the original wording: the models received the topic of each article, not a finished text to rewrite.
Graphite then compared the human-written and generated samples. The researchers looked at how often particular words and phrases appeared and how sentences were structured. The figures show differences between the groups of texts in this comparison. They do not mean that any phrase identified belongs exclusively to AI or can establish the origin of a specific publication.
The study compared not only different models but also versions of the same model. That matters when interpreting the findings: a tendency evident in one version may weaken in the next, while another takes its place. Graphite estimates that the overall number of characteristic features remains largely stable.
Words, Contrasts, and Em Dashes
For Claude Opus 5.5, the most distinctive word was “reliable”: it appeared 23 times more often in the model’s samples than in human-written texts. Opus 5.5 began avoiding the “it’s not X, it’s Y” structure but retained a similar way of emphasizing a point: “more than X, it’s Y.”
The model also frequently stressed the importance of what it was saying. “This matters” appeared 116 times more often in its samples than in the human-written examples, while “why X matters” appeared 92 times more often. These figures refer to the frequency of specific phrases in the samples studied, not to how useful the resulting text was.
The researchers identified a different set of constructions in OpenAI’s Astra. It more often introduced a qualification with “not just X” or contrasted approaches with “instead of relying on X.” By Graphite’s count, such wording appeared in Astra’s samples more than 100 times as often as in the human-written comparison texts. The model also tended to speak of “another dimension” of a topic and describe potential benefits with “can provide,” without giving a direct answer.
Em dash frequency also varied between models and versions. In Claude Opus 5.5 samples, the mark appeared 99% less often than in Opus 5 samples. Astra used it 88% less often than people in the control group, while Gemini 3.1 Pro almost stopped using it. Graphite documents these changes, but the figures alone cannot establish why they happened.
Graphite’s chief AI specialist, Greg Druck, reported a difference in word distribution: in his assessment, Claude models are moving closer to human-written texts over time, while GPT models are moving further away. This observation concerns word distribution, not the overall quality of their responses.
When Anthropic released Opus 5.5, it said the model communicated more naturally than earlier versions. When OpenAI released GPT-6-based Sol and Luna, it likewise said they were clearer and used less jargon and fewer unusual turns of phrase. These claims are not the same as the results of a frequency comparison: text that is easy to understand can still contain recurring constructions. Druck doubts that labs will be able to eliminate these features entirely; in his view, their ability to test large models is limited.
How to Apply the Findings to Website Content
I would use Graphite’s findings as a reason to reread AI-assisted content, not as a list of banned words. If several pages repeat the same explanations of why a topic matters or use the same contrast structure, it is worth checking whether those passages offer readers anything concrete. That is an editorial application of the findings, not a recommendation from the researchers themselves: Graphite compared texts but did not study website quality or propose publication rules.
My View
I would not try to identify an author by an em dash or a single recognizable phrase. In the samples studied, em dash frequency varied substantially even between versions of the same model. Its absence, in my view, proves nothing.
When editing, I find it more useful to look for repetition and check what each paragraph actually says. You can remove a typical turn of phrase and leave the same unclear thought behind. So I would sort out the meaning first, then work on the wording.
Sources
Where the news comes from. The text is a retelling in the author’s own words; the facts come from the source, the opinion is the author’s.
front-end development student
I’m learning to build websites and putting together small pages for my portfolio. Right now I’m figuring out responsive design and trying not to copy a solution until I understand why it works.
All posts by the authorRelated articles
Google Updates Its AI Content Guidelines: What to Check Before Publishing
Google has updated its AI content guidelines: texts need fact-checking and editing before publication. What does this mean for editorial teams and website owners?
Google Explains More Frequent Spam Updates as Content Volume Grows
Google says it is releasing spam updates more often because there is more content to assess and AI helps detect spam. Site owners should look at the quality of individual pages.
Google Warns Site Owners About Fake Content Authors
Google has added a warning about fake authors to its guidance for site owners. Editorial teams should check author names, photos, and credentials for accuracy.
Discussion 5
Anna Vogel
The strongest practical point is that this list can guide an edit, but it can’t tell you who wrote a piece; I’d treat repeated phrases as prompts to review, not proof of authorship.
Uliana Umanets
exactly — a phrase list is an editing checklist, not a tiny detective with a badge.
Maryna Koval
The detective-with-a-badge image works 😄 Especially since the study found model-specific phrases that changed between versions—useful clues for an edit, but far too unstable to pin authorship on.
João M.
One extra wrinkle: the study compared generated pieces with human articles after giving models summaries, so the phrases are patterns in that setup—not a universal checklist for every kind of web copy. In my edits, I’d also check whether a phrase fits the site’s voice and audience before changing it; otherwise we might sand off a perfectly natural sentence just because it looks familiar.
Beatriz Inês
I’d be careful about making “fits the site’s voice” the deciding filter: familiar phrasing can still be a useful signal even when it sounds on-brand. The study’s setup does limit what the list can tell us, but that’s a reason to use it as an editing prompt—not to wave away patterns that deserve a closer look.