Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)
TechCrunch Russell Brandom
A new study says AI writing still has giveaways, even after the obvious ones got fixed. The models swap one tell for another, so the fingerprints don’t really go away.
Based on reporting by TechCrunch, Russell Brandom — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AI prose has gotten better at hiding the old neon signs. The em dash has been pushed out. “Delve” is mostly gone. But a new study from the marketing firm Graphite says the models still leave plenty of tracks behind, just in different places every time the model changes.
Graphite built its analysis around 10,000 articles published before ChatGPT existed, then had AI models rewrite those pieces from summaries to keep source bias down. That let the firm compare human writing with model output at scale. The result was a long list of habits: 13,000 phrases that appeared at least twice as often in AI text as in human text. Greg Druck, Graphite’s chief AI officer, said Claude models are moving closer to human word distribution over time, while GPT models are moving further away.
Claude Opus 5.5, for instance, leans hard on the word “dependable,” which Graphite says shows up 23 times more often than in human samples. It has also moved away from the old “it’s not X, it’s Y” construction, but still reaches for a softer variation: something “is more than an X, it’s a Y.” The model’s favorite habit, though, is moral narration. “This matters” appears 116 times more often than in human writing, and “why X matters” is up 92 times.
OpenAI’s Astra has its own quirks. It likes to talk about “another dimension,” and it hedges with phrases like “may provide” and “can provide.” Graphite says its biggest giveaway is what it calls corrective framing, where a subject gets defined as “not simply X” or described as an alternative “rather than relying on X.” Those constructions were more than 100 times more common in Astra’s prose than in human writing.
The one tell everyone seems to be trying to kill is the em dash. Graphite says Opus 5.5 used it 99% less often than Opus, Astra is down 88% versus human samples, and Gemini 3.1 Pro has almost eliminated it. But Druck’s point is that the overall set of tells isn’t shrinking so much as reshuffling. The labs can sand off the most famous habits. The next layer shows up anyway.
My take — AI-written commentary, not fact-checked reporting
The obvious tells are getting cleaned up, which is exactly why this arms race is so tedious. AI writing is starting to look less like a robot and more like a committee memo that learned to smile. The real story isn’t that models can write naturally; it’s that they still can’t help announcing themselves, just in newer and more annoying ways.
Read more about this at: TechCrunch
