Graphite finds Opus 5.5 says “this is important” 116 times more

How Graphite looked for AI patterns

Graphite started with 10,000 articles published before ChatGPT’s release, then had different models rewrite them from short summaries. That approach was meant to reduce the influence of the original wording. Researchers compared the human and model samples for both word choices and broader sentence patterns.

Graphite defined a tell as an expression appearing at least twice as often in AI writing as in human writing. It found 13,000 such expressions.

Each model has its own tells

Claude Opus 5.5’s strongest tell is “reliable,” which appears 23 times more often in its sample than in human writing. It now avoids the “not X, but Y” construction, but often reaches for another: “it’s more than X, it’s Y.”

Its habit of spelling out significance is even more pronounced. “This is important” appears 116 times more often than in human writing, and “why X matters” appears 92 times more often.

OpenAI Astra has a different set of tics:

The dash is fading; the tells remain

Graphite’s sample suggests developers have responded to complaints about overused em dashes. Opus 5.5 used them 99% less often than Opus 5. Astra used them 88% less often than people did, and Gemini 3.1 Pro had almost stopped using them.

But the overall number of telltale patterns has mostly stayed the same, Graphite estimates. Druk told TechCrunch that models learn to drop their most obvious verbal habits, only for others to surface, with each version developing its own.

That persistence sits awkwardly beside companies’ claims that their models sound more human. When Anthropic released Opus 5.5, it said the model communicated more naturally than earlier versions; early users found its writing clearer and easier to read. OpenAI made similar claims for GPT-6 Sol and Luna, promising clearer prose, less jargon and fewer awkward turns of phrase.

I think the more interesting question is not whether labs can eliminate a particular phrase, but how well they can detect the next one. Druk doubts developers can remove these patterns entirely. His suggestion is that labs have less control over them than they assume: with models containing billions of parameters, they can test only a limited number of possibilities, leaving some habits unnoticed.

The dash may be easy to spot and suppress. The harder problem is that polishing away one recognizable tic does not necessarily make a model’s writing less recognizable.