Graphite's Opus 5.5 study, covered by TechCrunch on October 1, finds 2,548 words, phrases and sentence frames that the model uses disproportionately often in its test corpus.
The researchers compared articles on 9,974 aligned topics. Human reference articles predate ChatGPT; models produced rewrites from summaries so that the comparisons covered similar subjects. The count includes expressions used at least twice the human rate and meeting the study’s frequency filters.
Opus 5.5 uses “this matters” 116 times as often as the human reference texts. Other patterns repeatedly rate an idea’s importance or frame it through a contrast. The primary study reports fewer such tells than in Opus 5, but the reduction is small.
Frequency patterns can help explain why a batch of generated copy feels repetitive. They cannot reliably determine the author of one article. Graphite also separates word counts from its model-judged prose ratings, which remain judgments rather than direct measurements of human authorship.