Now that large language models (LLMs) are ubiquitous, humans are keen to spot the difference between human and AI-generated text. A new study by the marketing firm Graphite has identified over 13,000 phrases that AI content uses more frequently than human writing, revealing the subtle habits that give AI writing away. For instance, Claude Opus 5.5 frequently uses the phrase “dependable,” appearing 23 times more often than in human samples. Moreover, Opus 5.5 loves to emphasize why things matter, using “this matters” 116 times more than human writing. Meanwhile, OpenAI’s Astra prefers to describe “another dimension” and tends to hedge claims, using phrases like “may provide” or “can provide” more than 100 times more often than humans. While models like Opus 5.5 have reduced their reliance on em-dashes, other telltale phrases and structures persist.

The persistence of these phrases suggests that, despite significant advancements, AI models still exhibit human-like writing flaws. Graphite’s chief AI officer, Greg Druck, notes that while Claude models are getting closer to human word distribution, GPT models are moving further away. This indicates that the labs are making progress but are still grappling with controlling subtle writing habits.

The study’s findings highlight the ongoing challenge of creating AI that can mimic human writing without leaving telltale signs. OpenAI, for example, boasted that the GPT-6 versions of Sol and Luna would be clearer and less jargon-heavy. However, the persistence of these unique phrases and the fact that labs like Anthropic and Graphite are still refining their models suggest that perfecting human-like writing is an ongoing process. Druck’s skepticism about the labs’ ability to fully eliminate these telltale constructions is a reminder that the journey towards perfect AI writing is far from over.