Marketing firm Graphite has found that frontier AI models continue to rely on distinctive phrases and sentence patterns, identifying 13,000 phrases that appear at least twice as often in AI writing as in human prose.
Researchers used 10,000 articles published before ChatGPT’s release as a human control group, then had various models rewrite those articles from summaries to compare word choice and sentence construction.
Anthropic’s Claude Opus 5.5 favors the word “dependable” and frequently explains why things matter, using such phrasing up to 116 times more often than human writers. OpenAI’s Astra leans on hedges about what an action may or can provide and on corrective framing that defines a topic by what it is not, a construction more than 100 times more common than in human text.
All the labs have sharply cut em-dash use, with Gemini 3.1 Pro almost eliminating it. Yet Graphite chief AI officer Greg Druck said the overall number of tells is holding steady, as removing well-known habits allows new ones to emerge with each model version. He said Claude models are moving closer to human word patterns over time while GPT models are drifting further away, and suggested labs have less control over these quirks than expected given the scale of their models.
