How to spot AI-generated text

Since large language models work by predicting the next word in a sentence, they are more likely to use common words such as “the”, “he”, or “is” instead of odd and rare words. This is exactly the kind of text that automated detection systems are good at detecting, Ippolito and a team of Google researchers found in research they published in 2019.

But Ippolito’s study also showed something interesting: human participants tended to think that this type of “clean” text looked better and contained fewer errors, and therefore must have been written by a person.

In reality, human-written text is riddled with typos and is incredibly variable, incorporating different styles and slangs, while “language models very, very rarely mistype. They are much better at generating perfect texts,” says Ippolito.

“A typo in the text is actually a very good indicator that it was written by a human,” she adds.

The large language models themselves can also be used to detect AI-generated text. One of the most effective ways to do this is to retrain the model on some texts written by humans and others created by machines, so that it learns to differentiate between the two, says Muhammad Abdul-Mageed, incumbent Canada Research Chair in Natural Sciences. -language processing and machine learning at the University of British Columbia and studied detection.

Meanwhile, Scott Aaronson, a computer scientist at the University of Texas seconded as a researcher to OpenAI for a year, has developed watermarks for longer chunks of text generated by patterns such as GPT-3 – “a secret signal otherwise imperceptible in its choice of words, which you can use to prove later that, yes, it is from GPT,” he wrote in his blog.

An OpenAI spokesperson confirmed that the company is working on watermarks and said its policies state that users must clearly mark AI-generated text “in a way that no one could reasonably miss or misunderstand. “.

But these tech fixes come with some big caveats. Most of them don’t stand a chance against the latest generation of AI language models because 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 effective in some real-world use cases, like chatbots or messaging 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, says Abdul-Mageed.