Research by the marketing firm Graphite has revealed that artificial intelligence models continue to exhibit patterns that set their text apart from human writing. By comparing human and artificial intelligence texts, researchers identified 13,000 expressions that appeared at least twice as often in artificial intelligence content. According to Graphite’s chief artificial intelligence officer, Greg Druck, Claude models gradually move closer to the distribution of words used by humans, while GPT models are moving further away from that distribution.
The study used 10,000 articles published before ChatGPT was released as a human-generated control group. Researchers had different artificial intelligence models rewrite these articles based on their abstracts and compared their use of words, phrases and sentence structures.
- The most prominent hallmark of Claude Opus 5.5 is the word “dependable,” which it uses 23 times more often than human texts. The model also uses the phrase “this matters” 116 times more frequently and the pattern “why X matters” 92 times more frequently.
- OpenAI’s Astra model gravitates toward the phrase “another dimension” and patterns such as “may provide” or “can provide,” which indicate that an action could be beneficial. Structures described as “corrective framing,” including “not simply X” and “rather than relying on X,” appear more than 100 times as often in Astra’s texts as in human writing.
Models have reduced their excessive use of em dashes. Opus 5.5 uses this punctuation mark 99% less often than Opus 5, while Astra shows 88% lower usage than human samples; Gemini 3.1 Pro, meanwhile, has almost completely abandoned the em dash. Nevertheless, Graphite notes that the total number of AI-specific markers has not decreased, but that new patterns have merely replaced familiar ones. Druck argues that fully controlling such features is difficult in models with billions of parameters.
The research, published on October 1, 2026, shows that despite Anthropic and OpenAI’s claims of more natural, clear, and comprehensible writing, the models retain their distinctive patterns.
Why it matters
The findings show that identifying AI-generated text cannot be tied to a single marker, particularly easily noticeable uses such as em dashes. Because models can use other expressions more frequently while reducing a particular pattern, the criteria used in text evaluations need to change over time. This creates a methodological problem for readers and publishers who consider it important to distinguish content prepared with AI assistance from human writing. The study’s reliance on comparisons between human texts and model outputs suggests that the features identified are associated with specific generation tendencies rather than random preferences. However, it remains unclear how the gap between Anthropic and OpenAI’s goal of more natural writing and the models’ retention of distinctive patterns will be closed.