Now that LLM-generated prose is in every single place, human beings are keen for methods to smell it out. Whereas early tells like em-dashes and “delve” are lengthy gone, researchers say there are nonetheless loads of telltale habits that AI fashions fall again on when writing prose.
A new study from the advertising and marketing agency Graphite appeared on the writing habits of frontier fashions, sussing out every mannequin’s favourite phrases and phrases. Whereas previous tells like em-dash use have been stamped out, fashions nonetheless fall again on contrast-heavy constructions, with every mannequin model displaying its personal distinctive quirks. The largest shock is how broad the scope of tells seems to be. Graphite discovered 13,000 phrases that have been no less than twice as widespread within the AI content material as human content material — their definition of a “inform.”
“It seems that Claude fashions are literally getting nearer to the human phrase distribution over time,” Graphite’s chief AI officer Greg Druck informed TechCrunch. “And for the GPT fashions, it’s getting additional away.”
Learning AI-generated writing at scale required a cautious examine design. Graphite began with a corpus of 10,000 articles revealed earlier than the discharge of ChatGPT, serving because the human-generated management group. Then researchers had completely different AI fashions rewrite the articles from summaries, hoping to remove as a lot supply bias as doable. With matching samples from each people and every mannequin, they might examine how usually sure phrases and phrases appeared in AI writing, in addition to broader patterns in sentence development.
In line with Graphite’s outcomes, Claude Opus 5.5’s greatest inform is the phrase “reliable,” which pops up 23 instances extra usually than in human samples. Whereas Opus 5.5 now avoids the “it’s not X, it’s Y” sentence development, it nonetheless tends to say one thing “is greater than an X, it’s a Y.”
Above all, Opus likes to inform you why issues matter, utilizing the phrase “this issues” 116 instances extra usually than human writing, whereas “why X issues” happens 92 instances extra usually.
OpenAI’s Astra has a distinct set of tip-offs. This mannequin loves to explain “one other dimension” of no matter it’s speaking about, and tends to hedge claims by saying an motion “might present” or “can present” a specific profit. Its greatest inform is what Graphite calls the “corrective framing,” the place a subject is outlined as “not merely X” or supplied in its place, “somewhat than counting on X.” In line with graphite’s analysis, these constructions have been greater than 100 instances extra widespread in Astra-generated prose than in human writing.
Notably, all of the frontier labs appear to have responded to the concept that fashions overuse em-dashes. In Graphite’s samples, Opus 5.5 used the punctuation mark 99% much less usually than Opus 5. Astra now makes use of it 88% lower than human samples, whereas Gemini 3.1 Professional has nearly utterly eradicated the em-dash from its writing.
However whereas particular person tells change, Graphite says the general quantity is generally holding regular. “It’s not just like the tells are lowering,” Druck informed TechCrunch. “They’re managing to take away the most well-known tells, however different ones pop up. And each mannequin model has its personal.”
It’s stunning that tells are so persistent, given the labs’ deal with human-like writing kinds. In the Opus 5.5 release, Anthropic boasted that the mannequin “communicates extra naturally than prior fashions,” saying early customers “discovered its writing clearer and simpler to observe.”
OpenAI made comparable claims when releasing the GPT-6 versions of Sol and Luna, saying customers may “anticipate to see extra readability, much less jargon, [and] fewer odd turns of phrase.”
However Druck is skeptical about how a lot the labs can do to utterly remove telltale development or phrases.
“A basic speculation I’ve is that the labs are much less capable of management a few of these issues than you would possibly anticipate,” Druck says. “These are big fashions with billions of parameters. They’ve some finite variety of checks they will run, and issues slip by.”
Once you buy by hyperlinks in our articles, we might earn a small fee. This doesn’t have an effect on our editorial independence.