🔍 Intelligence Synthesis JP/EN
Ethics · Regulation · Technology — Pharma Practice Notes

AI Reports · MIT · 2026-11-18 · Blog / commentary

Helping AI text detectors keep up with changing conditions

AI text detectors can fail when AI-generated text is disguised as human writing, and they can also miss AI text as writing habits and new LLMs change. The researchers showed that test-time adaptation, which lets a deployed detector learn from new unlabeled data by adjusting its internal workings without retraining from scratch, makes detectors more accurate and resilient. In experiments, a commercial detector identified only about 27 percent of disguised AI-generated text.

Read the original ↗

All AI reports →