Helping AI text detectors keep up with changing conditions
Helping AI text detectors keep up with changing conditions
MIT and Cornell researchers showed that AI text detectors can be fooled by disguised AI-generated text and by natural shifts in writing over time, and proposed using test-time adaptation, which lets a deployed detector adjust to new unlabeled data without retraining. In experiments, a commercial detector identified about 27 percent of disguised AI text while the TTA approach reached 93 percent accuracy, and under simulated natural language shift TTA was 85 percent accurate versus 73 percent for traditional methods. The researchers caution TTA may not always work as future LLMs produce more natural-sounding language and verifiable human text becomes harder to find.
Source: MIT News AI · news.mit.edu