Researchers from MIT and Cornell University have proposed a test-time adaptation method that enables deployed AI text detectors to adjust their models using unlabeled new data, helping them cope with changes in writing styles and deliberate obfuscation. In experiments simulating natural variations in language, the method achieved 85% accuracy, compared with 73% for conventional methods. When detecting AI-generated content disguised as human-written text, its accuracy exceeded 93%, while existing methods achieved approximately 27%. The research will be presented at the Conference on Neural Information Processing Systems.

Source:MIT 研究新闻 · Read Original
Leave a Reply