Anyone who wants to know whether a text was written by AI tends to look for the models’ favorite words. A study by the marketing company Graphite now shows how unreliable that method is: the researchers found nearly 13,000 expressions that AI models use at least twice as often as humans, but the list shifts with every new model version. That is a useful insight for readers, teachers and newsrooms, precisely because it undercuts simple detection rules.
Key takeaways
- Graphite compared about 10,000 web articles from before ChatGPT with versions that nine AI models wrote on the same topics, and found 12,877 so-called tells.
- 65 percent of these features belong to only one model family, and even between two consecutive versions most of them differ.
- Claude Opus 5.5 has almost dropped the em dash but writes “this matters” 116 times as often as a human.
- According to the study, Claude’s word choice is moving closer to human writing, while GPT’s is moving away from it.
- The authors themselves advise against judging a text’s authorship based on individual expressions.
How the study works
The first version of the study came out on September 16, and an update adding Claude Opus 5.5 followed on October 1. The human comparison group was about 10,000 articles from the Common Crawl web archive, all published before ChatGPT launched on November 30, 2022. To keep the models from simply copying, the team first had GPT-4.1 summarize each article. Nine models from OpenAI, Anthropic and Google then wrote a new text of roughly the same length from that summary, including GPT-6 Astra, Claude Opus 5 and Gemini 3.1 Pro.
A “tell” is any word, phrase or sentence template with gaps that a model uses at least twice as often as the human authors. Graphite also measured stylistic features such as the em dash and had the texts rated for how mannered they read. That rating was done by a model, Claude Opus 5, which was not told who had written each text.
Every model has its own quirks
The results read like character studies. Claude Opus 5.5, which Anthropic released on September 22, uses the adjective “dependable” about 23 times as often as humans. Above all, it likes to stress that something is important: “this matters” appears 116 times as often, “why … matters” 92 times. The model has largely dropped the notorious “it’s not X, it’s Y” construction but instead writes along the lines of “is more than a …, it …” 98 times as often as humans. Opus 5.5 hardly uses the em dash anymore: 0.015 times per 1,000 words, compared with 2.92 for its predecessor Opus 5. That is a 99 percent drop.
OpenAI’s GPT-6 Astra has different habits. According to TechCrunch, it likes to describe “another dimension” of a topic and hedges with “may provide” or “can provide.” Its strongest tell is corrective framing such as “not simply X” or “rather than relying on X,” which appears more than 100 times as often as in human writing. Astra uses the em dash only about one-eighth as often as human authors. Gemini 3.1 Pro has nearly eliminated it but reaches for formal transitions instead: “furthermore” shows up 43 times as often as in human texts.
The reverse view is revealing, too. According to the study, humans use more than 100 times as many exclamation marks as each of the three current models and include far more parenthetical asides. So the absence of such quirks can be a clue as well.
The tells move instead of disappearing
The core finding is about change over time. Between the two newest versions of a model family, 55 to 72 percent of the tells belong to only one of the two versions. The best-known features, such as the word “delve” or formulaic closing lines, are becoming rarer, by 53 percent for Claude since Opus 4. But the total count is falling only slowly: Opus 4 had 3,746 tells, Opus 5.5 still has 2,548. For GPT, the number actually rose by 48 percent across the versions studied.
Graphite’s chief AI officer Greg Druck summed it up for TechCrunch: the labs are managing to remove the best-known tells, but other ones pop up, and every model version has its own. He suspects the companies are less able to control these quirks than one might expect. One figure supports a second observation: Claude’s word distribution gets closer to human texts with each version, and the gap shrank by 19 percent from Opus 5 to Opus 5.5. For GPT, it grew by 8 percent from GPT-5.6 Sol to Astra.
What this means for non-English texts
There is an important caveat for readers outside the English-speaking world. All the expressions measured are English, and the study did not examine whether the patterns carry over to other languages. When German reports turn “dependable” into “zuverlässig,” that is a plausible translation, not a measured result. The authors name further limits: only web articles written with a single fixed prompt were studied, and other instructions would produce other tells. The human comparison texts are also nearly four years old or older, so some supposed AI tells could simply reflect changes in how people write.
In practice, anyone checking a text should treat individual pet words at most as a reason to look more closely. A human can write “this is crucial,” and a model may stop doing so after its next update. Technical methods such as watermarks, which the provider embeds while generating the text, are more reliable. How far they go is shown by textGrain, which OpenAI is rolling out to mark ChatGPT texts in the EU, although the matching detector is not publicly available for now. There, too, a missing signal does not prove that a human wrote the text.
The list is useful, however, for anyone who writes with AI assistance. It shows which expressions should stand out when editing: exaggerated signals of importance, stilted contrasts, an excess of evaluative adjectives. Cutting such passages improves the text, whether or not anyone is looking for AI.
Outlook: style becomes a competitive arena
The providers now explicitly advertise more natural writing. At the launch of Opus 5.5, Anthropic promised the model communicates more naturally than its predecessors, and OpenAI held out the prospect of fewer odd turns of phrase with its new GPT-6 variants. Graphite’s measurements show that such promises can be checked, and that progress on one measure does not guarantee progress on another. For the question of who wrote a text, that leads to a sober rule: stylistic features can raise a suspicion, but they cannot settle it.