How to spot AI-generated text
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 rare words, complex. This is exactly the kind of text that automatic detection systems are good at collecting, Ippolito and a team of researchers at Google establish in the study they published in 2019.
But Ippolito’s study also found something interesting: human participants tended to think that this kind of “clean” text looked better and had fewer errors, and therefore had to be written by someone else. People.
In fact, human-written text has a lot of typos and is incredibly diverse, incorporating different styles and slang, while “language models very, very rarely make typos.” . They are much better at making perfect texts,” says Ippolito.
She added: “A typo in text is actually a really good sign that it was written by a human.
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 human-written and other machine-generated texts, so that it learns to distinguish between the two, says Muhammad Abdul -Mageed, Canada’s president of natural research, said. -language processing and machine learning at the University of British Columbia and has discovery research.
Meanwhile, Scott Aaronson, a computer scientist at the University of Texas, has been seconded as a researcher at OpenAI for a year. watermark development for longer passages of text produced by models like GPT-3—“another unnoticeable secret signal in word choices that you can later use to prove that, yes , this comes from GPT,” he writes on his blog.
An OpenAI spokesperson confirmed that the company is working on watermarks and said its policy stipulates that users should specify AI-generated text “in a way that no one can miss or understand.” reasonably erroneous”.
But these technical fixes come with big caveats. Most of them stand no chance against the latest generation of AI language models, as they are built on top of GPT-2 or older models. Many of these detection tools work best when there is a lot of text; they will be less effective in some specific use cases, such as chatbots or email assistants, that rely on shorter conversations and provide less data for analysis. And using large language models for detection also requires powerful computers and access to the AI model itself, something tech companies don’t allow, Abdul-Mageed said.