For example, since OpenAI’s chatbot ChatGPT launched in November, students are already using it to write essays and start cheating. The news website CNET uses ChatGPT to write articles, but only needs to issue a correction amid plagiarism accusations. However, there are promising ways to identify AI text. It’s about embedding hidden patterns that allow AI-generated text to be identified in these systems before it’s made public.
Studies have already shown that these watermarks can almost certainly identify AI-generated text. Developed by a team at the University of Maryland, it was able to find text created by Meta’s open source language model OPT-6.7B using a detection algorithm they built. The work is described in a paper that has yet to be peer-reviewed, and the code will be available for free around his February 15th.
AI language models work by predicting and generating one word at a time. After each word, the watermarking algorithm randomly splits the language model’s vocabulary into ‘greenlist’ and ‘redlist’ words, prompting the language model to choose a greenlist word.
The more greenlisted words in a passage, the more likely the text was machine-generated. Texts written by humans tend to contain more random word combinations. For example, for the word “beautiful”, a watermarking algorithm could classify the word “flower” as green and “orchid” as red. AI models using watermarking algorithms are more likely to use the word “flower” than “orchid,” explains Tom He Goldstein, an assistant professor at the University of Maryland who was involved in the study. .
