AI Detectors Keep Getting Fooled. Researchers Have a Fix.
Original headline: Helping AI text detectors keep up with changing conditions — news.mit.edu
More in AI and machine learning news
AI text detectors, used by educators and publishers to spot machine-generated writing, are surprisingly easy to fool. Researchers from MIT and Cornell demonstrated that a simple disguise can trick a commercial detector into missing over 70% of AI-generated text. The detectors also fail to adapt to natural shifts in writing style or new AI models.
To solve this, the team developed a training technique called test-time adaptation. This allows a deployed detector to learn from new, unlabeled data on the fly, adjusting its inner workings without a full retrain. In tests, this method boosted detection accuracy from a dismal 27% to a robust 93% against disguised text. This breakthrough could make detectors far more reliable for high-stakes uses in education, publishing, and cybersecurity, finally giving them a fighting chance against ever-evolving AI.
Read the original at news.mit.edu
Satirical commentary on real tech news, summarised with AI. Read the original source for the facts. How ZeroDaily works
