Graphite study exposes writing tells of Claude Opus 5.5

A new study by marketing firm Graphite has exposed the persistent linguistic quirks, or "tells," that continue to characterize text generated by frontier artificial intelligence models. By analyzing how different models rewrote summaries of 10,000 pre-ChatGPT articles, researchers identified 13,000 phrases that were at least twice as common in AI-generated content compared to human writing. While developers have successfully trained models to avoid older tells like em-dashes, new repetitive patterns have quickly emerged to take their place.

According to the findings, Anthropic's Claude Opus 5.5 heavily overuses the word "dependable," which appears 23 times more frequently than in human-written text. The model also frequently relies on the phrase "this matters," using it 116 times more often than humans, and "why X matters," which occurs 92 times more often. Meanwhile, OpenAI's Astra model favors hedging language like "may provide" or "can provide," alongside corrective framings such as "not simply X" or "rather than relying on X," which appeared over 100 times more often than in human samples.

The study highlighted how punctuation habits have shifted. Claude Opus 5.5 used em-dashes 99 percent less often than its predecessor, Opus 5, while Astra used them 88 percent less than humans, and Google's Gemini 3.1 Pro eliminated them almost entirely. Despite these adjustments, Graphite's chief AI officer Greg Druck noted that the overall volume of tells remains steady. While Claude models are gradually approaching human word distributions, OpenAI's GPT models are moving further away, even as releases like the GPT-6 versions of Sol and Luna promise clearer writing with less jargon.

For industry practitioners, these findings demonstrate that automated AI detection remains a moving target. Content creators and editors cannot rely on static lists of forbidden words, as each model update introduces a unique set of linguistic signatures. Because massive neural networks are difficult for labs to control perfectly, practitioners must continuously update their evaluation criteria to spot the latest stylistic tells of newly deployed models.