This tool is OpenAI’s response to the heat it’s getting from educators, journalists, and others to start ChatGPT with no way to identify generated text. However, it’s still very much a work in progress, and it’s unfortunately not very reliable. OpenAI claims that its AI text recognition correctly identifies 26% of AI-written text.
Obviously, OpenAI has a lot more work to do to refine the deviceThere is a limit to how well it can do. We are very unlikely to find a tool that can identify AI-generated text with 100% certainty. Prof. Mohamed Abdulmeged, who oversees research on Nature, said that since the main point of AI language models is to generate fluent and human-like text, and because the model mimics human-generated text, it is very difficult to recognize AI-generated text. Language Processing and Machine Learning at the University of British Columbia
Abdulmeged added. New AI language models are more powerful and better at generating more fluent language, making the current recognition toolkit obsolete.
OpenAI built the recognizer by creating a completely new AI language model similar to ChatGPT, which trains itself to distinguish outputs from similar models. Although details are sparse, the company trained the model on examples of AI-generated text and examples of human-generated text, and then asked it to view the AI-generated text. We asked for more information, but OpenAI did not respond.
Last month, I wrote about another technique for finding AI-generated text: watermarks. These act as a sort of secret symbol in AI-generated text for computer programs to recognize.
University of Maryland researchers have developed a neat way to apply watermarks to text generated by AI language models, and they’ve made it available for free. These watermarks allow us to tell with absolute certainty when AI-generated text has been used.
The problem is that this method requires AI companies to embed watermarking in their chatbots from the very beginning. OpenAI is developing these systems but has not yet deployed them in any of its products. Why is it late? One reason is that watermarking AI-generated text may not always be desirable.
One of the most promising ways ChatGPT can be integrated into products is as a tool to help you write emails or an improved spell checker in a word processor. That’s not exactly cheating. But watermarking all AI-generated text automatically flags these results and can lead to false accusations.
