A new study from marketing firm Graphite suggests that AI-generated prose remains detectable, even as older tells disappear. The researchers compared 10,000 pre-ChatGPT human articles with AI rewrites, looking for words and phrases that appeared at least twice as often in machine text. They found 13,000 such phrases, and each model had its own quirks.
Claude Opus 5.5, for example, used 'dependable' 23 times more often than humans and leaned on constructions like 'more than an X, it's a Y.' Its biggest signature was declaring that something 'matters,' with 'this matters' appearing 116 times more often than in human writing. OpenAI's Astra, meanwhile, favored 'another dimension' and hedged with 'may provide' or 'can provide,' while relying on corrective framing like 'not simply X.'
The labs have clearly tried to clean up one famous tell: em-dashes. Opus 5.5 used them 99% less than its predecessor, and Gemini 3.1 Pro has nearly eliminated them. But Graphite's chief AI officer, Greg Druck, said the overall number of tells is holding steady. He expressed skepticism that labs can fully control these habits, noting that models are too large and complex to test every quirk away. The findings suggest that AI writing may remain identifiable for the foreseeable future, even as the specific markers evolve.