Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)

Graphite analyzed rewrites produced by several frontier large language models against a human-generated control set of 10,000 articles published before ChatGPT, aiming to reduce source bias by generating model outputs from summaries. Comparing matched samples allowed researchers to measure how frequently specific words, phrases and constructions appeared in each model relative to human writing. The firm defined a “tell” as any phrase at least twice as common in AI outputs as in the human corpus and found roughly 13,000 such phrases across the models.

The study surfaces pronounced, model-specific habits. Anthropic’s Claude Opus 5.5 disproportionately uses the adjective “dependable,” which showed up 23 times more often than in human samples; it also favors constructions emphasizing importance, with the phrase “this matters” appearing 116 times more often and “why X matters” 92 times more often than in human writing. Opus 5.5 has reduced some previously common forms — Graphite reports the model now avoids the “it’s not X, it’s Y” template — but retains other patterned usages such as “is more than an X, it’s a Y.”

OpenAI’s Astra exhibits different patterns: it frequently frames topics as having “another dimension,” uses hedging verbs like “may provide” or “can provide,” and shows a high prevalence of what Graphite calls “corrective framing” (phrases positioning something as “not simply X” or “rather than relying on X”). According to the report, those corrective constructions were more than 100 times more common in Astra outputs than in the human baseline. At the same time, many labs appear to have addressed the em-dash overuse that earlier research flagged: Graphite found Opus 5.5 used em-dashes 99% less often than the human samples, Astra 88% less, and Google’s Gemini 3.1 Pro has almost entirely eliminated the mark from its samples.

Graphite’s chief AI officer Greg Druck told TechCrunch that Claude-family models are trending toward human-like word distributions while GPT-family models are moving further away. Druck also expressed skepticism that labs can completely eradicate tells, arguing that enormous models and practical testing limits mean some patterned behaviors will “slip through.” Anthropic and OpenAI have both promoted the improved naturalness and clarity of recent releases — Anthropic for Opus 5.5 and OpenAI for GPT-6 Sol and Luna — yet Graphite’s results indicate many detectable stylistic markers remain.