Ask a language model to rewrite an article and it will hand back prose that reads fluently — and, according to a new study, reliably gives itself away.
The marketing firm Graphite compared how frontier models rewrite text against human writing, using a corpus of 10,000 articles published before the release of ChatGPT as a control group and having different AI models rewrite those articles from summaries. The result: about 13,000 phrases that appear at least twice as often in AI-generated prose as in human prose — the study's definition of a "tell".
Each model has its own habits. Anthropic's Claude Opus 5.5 leans on the word "dependable", which appears 23 times more often than in the human samples, and is far more likely to tell readers why something matters: "this matters" shows up 116 times more often, and "why X matters" 92 times more often. According to Graphite, Opus 5.5 has largely shed the "it's not X, it's Y" construction but still favours "is more than an X, it's a Y".
OpenAI's Astra has a different fingerprint. It reaches for "another dimension" of a subject, hedges with "may provide" or "can provide", and is especially fond of what Graphite calls corrective framing — defining something as "not simply X", or offering an alternative, "rather than relying on X". Those constructions were more than 100 times more common than in human writing.
The familiar em-dash tell is fading. Opus 5.5 used the punctuation mark 99% less often than Opus 5, Astra uses it 88% less than the human baseline, and Gemini 3.1 Pro has all but eliminated it. Graphite's chief AI officer Greg Druck told TechCrunch that Claude models are drifting closer to the human word distribution over time while GPT models are moving further away, and that the overall number of tells is holding steady even as the best-known ones are stamped out. His working hypothesis: labs control less of this than they claim, because "these are giant models with billions of parameters" where things slip through.
