Opus 5.5 says 'this matters' 116x more than humans

The study's design matters for how much weight the findings can carry. Graphite used a corpus of 10,000 articles published before ChatGPT was released as a human control, then had each AI model rewrite those articles from summaries to reduce source bias. That setup lets the researchers compare the same underlying content written by humans and by each model, rather than comparing unrelated texts. The 13,000-phrase figure is the study's threshold for a tell — a phrase at least twice as common in AI content as in human content — and it is unusually broad compared with earlier, more anecdotal lists of AI giveaways.

The findings also show that tells shift rather than shrink. All the frontier labs appear to have responded to the criticism that models overuse em-dashes: Opus 5.5 used them 99% less than Opus 5, Astra 88% less than human samples, and Gemini 3.1 Pro has almost eliminated them. Yet Graphite's chief AI officer, Greg Druck, says the overall number of tells is mostly holding steady, with each model version displaying its own quirks. That pattern is notable because Anthropic and OpenAI both marketed their newest releases on more natural writing — Anthropic said Opus 5.5 "communicates more naturally than prior models," and OpenAI said users of the GPT-6 versions of Sol and Luna could expect "more clarity, less jargon, [and] fewer odd turns of phrase." Druck is skeptical that labs can fully control these habits, noting that giant models with billions of parameters have a finite number of tests they can run.

The practical stakes sit with anyone who has to judge whether a piece of prose was written by a person or a model — editors, publishers, recruiters, and admissions reviewers among them. The study suggests the old shortcuts, like scanning for em-dashes or the word "delve," are no longer reliable, since labs have largely stamped those out. What remains to be seen is whether the underlying rate of tells stays steady as models get larger, or whether more targeted fixes can bring individual quirks down without new ones appearing.