A team of researchers from MIT and Cornell University has developed a more accurate AI text detector that can adapt to changing writing habits and new AI models. The detector uses a method called test-time adaptation, which allows it to learn from new, unlabeled data and adjust its inner workings on the fly. This makes it more resilient to deception and able to identify AI-generated text that has been disguised to appear human-written. In experiments, the new approach achieved a 93% accuracy rate, compared to a commercial detector that could only identify about 27% of AI-generated text. The researchers believe that this method could be used in various high-stakes contexts, such as scientific publishing, short story competitions, and cybersecurity, to prevent email phishing scams and identify search engine optimization fraud. The team's goal is to make it more difficult for people to pass off AI-generated writing as human-written, and to help educators, editors, publishers, and hiring managers detect and prevent AI-generated content. The study's findings have significant implications for the development of more accurate AI detectors and the prevention of AI-generated content from being misattributed as human-written.