Detector that keeps learning catches far more disguised AI text
Ongoing: AI Hot

Researchers at MIT and Cornell Tech found that AI text detectors can be fooled by AI writing disguised as human. They tested a method that lets a deployed detector keep learning from new, unlabelled text.
The method is called test-time adaptation. Manish Raghavan of the MIT Sloan School of Management, Kevin Ren and Nikhil Garg of Cornell Tech wrote the paper, which they will present at the Conference on Neural Information Processing Systems. In their experiments, a commercial detector found about 27% of disguised AI text, while the new approach reached 93% accuracy. That figure comes from the researchers' own tests, and it may not hold elsewhere.