Former Council Member Melanie Mitchell on Detecting AI Generated Text

AI generative models have recently taken the world by storm creating digital art, music and prose. Among these innovations is a new language model called ChatGPT on openai.com. Using this tool, users can ask ChatGPT anything ranging from answering homework questions to identifying potential research directions. Something that sets ChatGPT apart from other language models is its ability to create more natural and accurate language when answering questions and requests. ChatGPT’s answers are extremely similar if not identical to human responses. This has presented a multitude of concerns over plagiarism and misinformation.

Something that could help curb these concerns is a means of detecting text generated by language models. Former CCC council member and Professor at the Santa Fe Institute, Melanie Mitchell wrote an article on her blog “AI: A Guide for Thinking Humans” addressing research directions and methods that could be used to detect AI generated text. The article, “On Detecting Whether Text was Generated by a Human or an AI Language Model” discusses the potential urgency of addressing this issue and breaks down potential technological solutions.

In the blog, Mitchell walks readers through the science behind large language models and potential means of detection. Citing two articles “DetectGPT Zero-Shot Machine-Generated Text Detection using Probability Curvature” and a “A Watermark for Large Language Models”, Mitchell discusses using probability and or watermarks as detection techniques and explains how and why such techniques may work.

You can read her full post here.

The Computing Community Consortium’s task force, NextGen AI, has been having discussions surrounding the opportunities and challenges that generative models present. Established in fall 2022, the task force focuses on avenues to broaden the impacts of AI by examining and overcoming the limits and pitfalls of the field.