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

Now that LLM-generated prose is everywhere, quality beings are anxious for ways to sniff it out. While aboriginal tells similar em-dashes and “delve” are agelong gone, researchers accidental determination are inactive plentifulness of telltale habits that AI models autumn backmost connected erstwhile penning prose.

A caller study from the selling steadfast Graphite looked astatine the penning habits of frontier models, sussing retired each model’s favourite words and phrases. While aged tells similar em-dash usage person been stamped out, models inactive autumn backmost connected contrast-heavy constructions, with each exemplary mentation showing its ain unsocial quirks. The biggest astonishment is however wide the scope of tells turns retired to be. Graphite recovered 13,000 phrases that were astatine slightest doubly arsenic communal successful the AI contented arsenic quality contented — their explanation of a “tell.” 

“It turns retired that Claude models are really getting person to the quality connection organisation implicit time,” Graphite’s main AI serviceman Greg Druck told TechCrunch. “And for the GPT models, it’s getting further away.”

Studying AI-generated penning astatine standard required a cautious survey design. Graphite started with a corpus of 10,000 articles published earlier the merchandise of ChatGPT, serving arsenic the human-generated power group. Then researchers had antithetic AI models rewrite the articles from summaries, hoping to destruct arsenic overmuch root bias arsenic possible. With matching samples from some humans and each model, they could comparison however often definite words and phrases appeared successful AI writing, arsenic good arsenic broader patterns successful condemnation construction.

According to Graphite’s results, Claude Opus 5.5’s biggest archer is the connection “dependable,” which pops up 23 times much often than successful quality samples. While Opus 5.5 present avoids the “it’s not X, it’s Y” condemnation construction, it inactive tends to accidental thing “is much than an X, it’s a Y.”

Above all, Opus loves to archer you wherefore things matter, utilizing the operation “this matters” 116 times much often than quality writing, portion “why X matters” occurs 92 times much often.

OpenAI’s Astra has a antithetic acceptable of tip-offs. This exemplary loves to picture “another dimension” of immoderate it’s talking about, and tends to hedge claims by saying an enactment “may provide” oregon “can provide” a peculiar benefit. Its biggest archer is what Graphite calls the “corrective framing,” wherever a taxable is defined arsenic “not simply X” oregon offered arsenic an alternative, “rather than relying connected X.” According to graphite’s research, those constructions were much than 100 times much communal successful Astra-generated prose than successful quality writing.

Notably, each the frontier labs look to person responded to the thought that models overuse em-dashes. In Graphite’s samples, Opus 5.5 utilized the punctuation people 99% little often than Opus 5. Astra present uses it 88 percent little than quality samples, whereas Gemini 3.1 Pro has astir wholly eliminated the em-dash from its writing.

But portion idiosyncratic tells change, Graphite says the wide fig is mostly holding steady. “It’s not similar the tells are decreasing,” Druck told TechCrunch. “They are managing to region the most well-known tells, but different ones popular up. And each exemplary mentation has its own.”

It’s astonishing that tells are truthful persistent, fixed the labs’ absorption connected human-like penning styles. In the Opus 5.5 release, Anthropic boasted that the exemplary “communicates much people than anterior models,” saying aboriginal users “found its penning clearer and easier to follow.”

OpenAI made akin claims erstwhile releasing the GPT-6 versions of Sol and Luna, saying users could “expect to spot much clarity, little jargon, [and] less unusual turns of phrase.”

But Druck is skeptical astir however overmuch the labs tin bash to wholly destruct telltale operation oregon phrases.

“A wide proposal I person is that the labs are little capable to power immoderate of these things than you mightiness expect,” Druck says. “These are elephantine models with billions of parameters. They person immoderate finite fig of tests they tin run, and things gaffe through.”