There’s a cat-and-mouse game between those using generative AI chatbots to produce text undetected and those trying to catch them. Many believe they know the telltale signs—though as a journalist fond of the word “delve” and prone to em-dashes, I’m not so sure.
Researchers at four U.S. universities, however, have taken a more rigorous approach, identifying linguistic fingerprints that reveal which large language model (LLM) produced a given text.
“All these chatbots are coming out every day, and we interact with them, but we don’t really understand the differences between them,” says Mingjie Sun, a researcher at Carnegie Mellon University and lead author of the study, which was published in Cornell University’s preprint server arXiv. “By training a machine learning classifier to do this task, and by looking at the performance of that classifier, we can then assess the difference between different LLMs.”
Sun and his colleagues developed a machine learning model that analyzed the outputs of five popular LLMs, and was able to distinguish between them with 97.1% accuracy. Their machine learning model uncovered distinct verbal quirks unique to each LLM.
ChatGPT’s GPT-4o model, for instance, tends to use “utilize” more than other models. DeepSeek is partial to saying “certainly.” Google’s Gemini often prefaces its conclusions with the word “essentially,” while Anthropic’s Claude overuses phrases like “according to” and “according to the text” when citing its sources.
xAI’s Grok stands out as more discursive and didactic, often reminding users to “remember” key points while guiding them through arguments with “not only” and “but also.”
“The writing, the word choices, the formatting are all different,” says Yida Yin, a researcher at the University of California, Berkeley, and a coauthor of the paper.
These insights can help users select the best model for specific writing tasks—or aid those trying to catch AI-generated text masquerading as human work. So, remember: according to this study, if a model utilizes certain words, it’s certainly possible to identify it.
Login to add comment
Other posts in this group
Rather than the Sunday scaries or toxic bosses, employees have unlocked a new workplace fear: office chair butt.
While not a new concern, the term has resurfaced on TikTok to describe ho
Music streaming service Deezer said Friday that it will start flagging albums with AI-generated songs, part of its fight against
In 2023, the flagship reveal at Apple’s WWDC keynote was unquestionably the debut of
What happens when venture capital and government pull back from science entrepreneurs at the same time? Many scientists think we’re about to find out, and are looking at how we can preserve our co
