A watermark for chatbots can spot text written by an AI

For example, since the launch of OpenAI’s ChatGPT chatbot in November, students have already started cheating by using it to write essays for them. News website CNET used ChatGPT to write articles, only to have to post corrections amid accusations of plagiarism. Building the watermark approach into such systems before they are released could help resolve these issues.

In studies, these watermarks have already been used to identify AI-generated text with near certainty. Researchers at the University of Maryland, for example, were able to spot text created by Meta’s open-source language model, OPT-6.7B, using a detection algorithm they built. The work is described in an article that has not yet been peer-reviewed, and the code will be freely available around February 15.

AI language models work by predicting and generating one word at a time. After each word, the watermarking algorithm randomly divides the language model’s vocabulary into words on a “green list” and a “red list”, then prompts the model to choose words from the green list.

The more whitelisted words in a passage, the more likely the text was machine-generated. Text written by one person tends to contain a more random mix of words. For example, for the word “beautiful”, the watermarking algorithm might classify the word “flower” as green and “orchid” as red. The AI ​​model with the watermark algorithm would more likely use the word “flower” than “orchid,” says Tom Goldstein, an assistant professor at the University of Maryland who was involved in the research.