Wondering If AI Wrote That Document? These Are the Surprising New Clues

Artificial intelligence has come a long way in a short time, but it still has quite some distance to go before its writing abilities are indistinguishable from human authors’. A new study from Graphite, a San Francisco-based marketing firm, has examined the writing styles of the major frontier AI models. And while you won’t be able to catch AI-written prose quite as easily as you used to (when the models couldn’t get enough of the word “delve” and had em dashes everywhere), there are still plenty of signs to look for. The good news for AI companies is the gap is closing, albeit slowly. An October 1 study found Anthropic’s Claude Opus 5.5 had 2,548 tells that a document was being written by AI. That’s down from 3,746 with Opus 4. An earlier study by the group, which was released in mid-September, said it collectively found almost 13,000 tells among all frontier AI models. An earlier model of Claude Opus (Opus 5), for example, used the pattern “less like a _ and more like” roughly 105 times as often as human writers. Graphite didn’t update the frequency count of that particular phrase in its updated study, but it did note that the overall word-distribution divergence (a statistical measure of how differently two texts use words overall) from human writing is 19 percent lower than it was with Opus 5, which was released on July 24. Not all AI models are improving, though. The word-distribution divergence of GPT-6 Astra, released September 4, is 8 percent higher than that of GPT-5.6 Sol, which debuted on July 9 of this year. In other words, Astra is getting worse at writing like us. So long, em dashes One message was clearly received by programmers, though—get rid of the em dash. Opus 5.5 uses em dashes 99 percent less often than Opus 5, the study found, while Astra uses them 95 percent less often than GPT-5. Both, in fact, use the punctuation less than human-written stories—like this one. Still, there are plenty of tells. The study of Opus 5.5 found that the AI continues to have some favorite words it defaults to, though maybe not quite as much as it did with “delve” in the early days. “Dependable” is a favorite of the Anthropic model, getting used 23 times more often than humans use the word. “Clearer” is also a high-ranker, being used 14 times as frequently. Other go-to words included “matters” (13X), “quietly,” “practical,” and “steady” (all 11X). When it comes to transitioning between ideas, Opus 5.5 is a tremendous fan of the phrase “looking ahead the,” using it 40 times more than the human rate. Other popular picks include “Adds another later” (27X), “What comes next (24X), and “in practice” (7X). Those tells pale in comparison with the ones AI exhibits when it tries to point out things it feels readers might miss. Opus 5.5 used “this matters” 116 times more frequently than humans—and “why __ matters” 92 times more often. And when describing something by contrasting it with another item or idea, it favored “is more than a __, it” (e.g., “alcohol is more than a disinfectant, it can also be consumed”), using the phrase 98 times more than the human rate. (“Rather than simply” was the runner up, with a 32X usage.) Researchers also compared Opus 5.5 with GPT-6 Astra and found that AI tells were hardly universal. Some 65 percent of tells are unique to one model family, the September study found. Astra, for instance, used the phrase “does not establish” 275 times more than Opus 5.5. Anthropic’s model, meanwhile, leaned into some superlatives 79 times more frequently than Astra. Despite it all, Graphite notes that Opus 5.5 is making progress toward mimicking human writing, even if it still has a long way to go. “Some tells become less common, while others become more common,” it wrote. “Opus 5.5 uses well-known tells less often and its overall word use is more similar to human writing, but it still disproportionately uses thousands of words, phrases, and frames.” BY CHRIS MORRIS @MORRISATLARGE