\n<\/p>\n
Now that LLM-generated prose is everywhere, human beings are eager for ways to sniff it out. While early tells like em-dashes and \u201cdelve\u201d are long gone, researchers say there are still plenty of telltale habits that AI models fall back on when writing prose.<\/p>\n
A new study<\/a> from the marketing firm Graphite looked at the writing habits of frontier models, sussing out each model\u2019s favorite words and phrases. While old tells like em-dash use have been stamped out, models still fall back on contrast-heavy constructions, with each model version showing its own unique quirks. The biggest surprise is how broad the scope of tells turns out to be. Graphite found 13,000 phrases that were at least twice as common in the AI content as human content \u2014 their definition of a \u201ctell.\u201d\u00a0<\/p>\n \u201cIt turns out that Claude models are actually getting closer to the human word distribution over time,\u201d Graphite\u2019s chief AI officer Greg Druck told TechCrunch. \u201cAnd for the GPT models, it\u2019s getting further away.\u201d<\/p>\n Studying AI-generated writing at scale required a careful study design. Graphite started with a corpus of 10,000 articles published before the release of ChatGPT, serving as the human-generated control group. Then researchers had different AI models rewrite the articles from summaries, hoping to eliminate as much source bias as possible. With matching samples from both humans and each model, they could compare how often certain words and phrases appeared in AI writing, as well as broader patterns in sentence construction.<\/p>\n According to Graphite\u2019s results, Claude Opus 5.5\u2019s biggest tell is the word \u201cdependable,\u201d which pops up 23 times more often than in human samples. While Opus 5.5 now avoids the \u201cit\u2019s not X, it\u2019s Y\u201d sentence construction<\/a>, it still tends to say something \u201cis more than an X, it\u2019s a Y.\u201d <\/p>\n Above all, Opus loves to tell you why things matter, using the phrase \u201cthis matters\u201d 116 times more often than human writing, while \u201cwhy X matters\u201d occurs 92 times more often.<\/p>\n OpenAI\u2019s Astra has a different set of tip-offs. This model loves to describe \u201canother dimension\u201d of whatever it\u2019s talking about, and tends to hedge claims by saying an action \u201cmay provide\u201d or \u201ccan provide\u201d a particular benefit. Its biggest tell is what Graphite calls the \u201ccorrective framing,\u201d where a topic is defined as \u201cnot simply X\u201d or offered as an alternative, \u201crather than relying on X.\u201d According to graphite\u2019s research, those constructions were more than 100 times more common in Astra-generated prose than in human writing.<\/p>\n Notably, all the frontier labs seem to have responded to the idea that models overuse em-dashes. In Graphite\u2019s samples, Opus 5.5 used the punctuation mark 99% less often than Opus 5. Astra now uses it 88% less than human samples, whereas Gemini 3.1 Pro has almost completely eliminated the em-dash from its writing.<\/p>\n But while individual tells change, Graphite says the overall number is mostly holding steady. \u201cIt\u2019s not like the tells are decreasing,\u201d Druck told TechCrunch. \u201cThey are managing to remove the\u00a0most well-known tells, but other ones pop up. And every model version has its own.\u201d<\/p>\n It\u2019s surprising that tells are so persistent, given the labs\u2019 focus on human-like writing styles. In the Opus 5.5 release<\/a>, Anthropic boasted that the model \u201ccommunicates more naturally than prior models,\u201d saying early users \u201cfound its writing clearer and easier to follow.\u201d<\/p>\n OpenAI made similar claims when releasing the GPT-6 versions of Sol and Luna<\/a>, saying users could \u201cexpect to see more clarity, less jargon, [and] fewer odd turns of phrase.\u201d<\/p>\n But Druck is skeptical about how much the labs can do to completely eliminate telltale construction or phrases.<\/p>\n \u201cA general hypothesis I have is that the labs are less able to control some of these things than you might expect,\u201d Druck says. \u201cThese are giant models with billions of parameters. They have some finite number of tests they can run, and things slip through.\u201d<\/p>\n<\/div>\n When you purchase through links in our articles, we may earn a small commission<\/a>. This doesn\u2019t affect our editorial independence.<\/em><\/p>\n