Because large language models work by predicting the next word in a sentence, they are more likely to use common words like “the,” “it,” or “is” instead of wonky, rare words. This is exactly the kind of text that automated detector systems are good at picking up, Ippolito and a team of researchers at Google found in research they published in 2019.
But Ippolito’s study also showed something interesting: the human participants tended to think this kind of “clean” text looked better and contained fewer mistakes, and thus that it must have been written by a person.
In reality, human-written text is riddled with typos and is incredibly variable, incorporating different styles and slang, while “language models very, very rarely make typos. They’re much better at generating perfect texts,” Ippolito says.
“A typo in the text is actually a really good indicator that it was human written,” she adds.
Large language models themselves can also be used to detect AI-generated text. One of the most successful ways to do this is to retrain the model on some texts written by humans, and others created by machines, so it learns to differentiate between the two, says Muhammad Abdul-Mageed, who is the Canada research chair in natural-language processing and machine learning at the University of British Columbia and has studied detection.
Scott Aaronson, a computer scientist at the University of Texas on secondment as a researcher at OpenAI for a year, meanwhile, has been developing watermarks for longer pieces of text generated by models such as GPT-3—“an otherwise unnoticeable secret signal in its choices of words, which you can use to prove later that, yes, this came from GPT,” he writes in his blog.
A spokesperson for OpenAI confirmed that the company is working on watermarks, and said its policies state that users should clearly indicate text generated by AI “in a way no one could reasonably miss or misunderstand.”
But these technical fixes come with big caveats. Most of them don’t stand a chance against the latest generation of AI language models, as they are built on GPT-2 or other earlier models. Many of these detection tools work best when there is a lot of text available; they will be less efficient in some concrete use cases, like chatbots or email assistants, which rely on shorter conversations and provide less data to analyze. And using large language models for detection also requires powerful computers, and access to the AI model itself, which tech companies don’t allow, Abdul-Mageed says.