Frontier language models still leave thousands of linguistic fingerprints despite years of tuning for natural style. Marketing firm Graphite compared model rewrites with 10,000 human articles and counted 13,000 phrases at least twice as common in AI prose. Claude Opus 5.5 uses "this matters" 116 times more often than people. For business, this changes how AI drafts can be checked before publication.
How Graphite measured model writing habits
Graphite built its comparison on 10,000 articles published before the release of ChatGPT as a human control group. Researchers then asked different models to rewrite those articles from summaries, a design meant to reduce source bias. With matched human and machine samples, the team counted word frequency and sentence patterns. A tell was defined as a phrase at least twice as common in AI content as in human writing. That threshold produced 13,000 such phrases across frontier models.
Each model version showed its own vocabulary skew. Opus 5.5 overuses "dependable" by 23 times versus humans and favors "is more than an X, it is a Y," while avoiding the older "it is not X, it is Y" form. OpenAI Astra leans on "another dimension" and hedging verbs such as "may provide" or "can provide" a benefit. Its strongest signal is corrective framing like "not simply X" and "rather than relying on X," over 100 times more common than in human prose. Gemini 3.1 Pro was included in the punctuation comparison as well.
The study lands after repeated industry claims about more natural output. Anthropic said Opus 5.5 communicates more naturally than prior models, with early users finding its writing clearer and easier to follow. OpenAI said GPT-6 versions of Sol and Luna would bring more clarity, less jargon and fewer odd turns of phrase. Graphite chief AI officer Greg Druck said Claude models are moving closer to human word distribution over time, while GPT models are moving further away. Old signals such as em-dash use and "delve" have largely faded from current output.
What persistent AI tells mean for business
Companies that publish AI-assisted drafts gain a practical quality filter from these lists. Editors can search for "this matters," "why X matters," "another dimension," and corrective frames before release, especially in marketing, help centers, and sales collateral. The method scales because it rests on frequency counts rather than subjective taste. Small teams get a fast first pass without extra software, while large content operations can encode the phrases into style checks and approval workflows.
The same signals carry limits that buyers should weigh. Graphite notes that well-known tells get removed while new ones appear, so any fixed list needs regular updates for each model version. Em-dash use illustrates the shift: Opus 5.5 uses it 99% less often than Opus 5, Astra uses it 88% less than human samples, and Gemini 3.1 Pro has almost eliminated it. Druck attributes the persistence to giant models with billions of parameters and finite testing, where some patterns slip through. A phrase match alone does not prove authorship or factual error.
A useful marker will be the next round of model updates and independent recounts of the same tells. If "dependable," "this matters," and "not simply X" fall sharply without replacement phrases rising, tuning will have narrowed the gap. If total tells stay near 13,000 with different leaders, detection will remain a moving target. Content teams should then treat style audits as recurring work tied to model versions.
