Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)
Researchers from Graphite, a marketing firm, have analysed writing patterns across frontier AI models and identified thousands of telltale phrases and constructions that distinguish AI-generated text from human writing. Despite AI companies claiming their latest models communicate more naturally, the study found that whilst old tells like em-dashes have been largely eliminated, new linguistic quirks have emerged to replace them, suggesting that AI writing remains distinctly identifiable.
The study examined how Claude Opus 5.5, OpenAI's Astra, and Gemini 3.1 Pro rewrite content, identifying 13,000 phrases appearing at least twice as often in AI text as in human writing. Claude Opus 5.5's most prominent tell is its excessive use of "this matters" (116 times more frequent) and "why X matters" (92 times more frequent), whilst Astra favours corrective framings such as "not simply X" and "rather than relying on X" (over 100 times more common). The research suggests that whilst individual tells change across model versions, the overall number of identifiable patterns remains steady.
- AI models retain identifiable quirks despite natural-sounding claims
- Claude Opus 5.5 overuses "this matters" 116 times more than humans
- Different models show distinct patterns that evolve but persist
New here? Start with this
Artificial intelligence language models have become increasingly common for generating text. AI companies claim that their latest models now communicate more naturally than earlier versions, whilst a study by Graphite, a marketing firm, has identified distinctive linguistic patterns in how these models generate text.
Researchers examined three leading AI models—Claude Opus 5.5, OpenAI's Astra, and Gemini 3.1 Pro—and found thousands of phrases that appear far more frequently in AI text than in human writing. Claude Opus 5.5, for instance, uses "this matters" over 100 times more frequently than humans do, whilst Astra employs corrective framings like "not simply X" at similarly elevated rates.
These patterns matter because they could help identify AI-generated content. Although older tells such as excessive em-dashes have been eliminated as models have improved, the research shows that new distinctive patterns have emerged to replace them.