An AI text detector that adapts finds 93% of disguised text
A detector from a company found only about 27% of disguised text, and the adapting method found more than 93%.
Researchers from MIT and Cornell University showed that tools that find AI text can make errors. This happens when the AI text is changed to look the same as text of persons. They then used test-time adaptation, a method that lets a detector learn from new text that has no labels after release. In their tests, the result on disguised text went from about 27% to more than 93%.
How it works
Makers train a detector on text that they know is AI or from persons. After release, users change how they write. Some users operate a humanizer, a tool that changes AI text to make it look the same as text of persons. The new method learns from the new text that users send. It uses no labels.
The numbers
- Disguised text: about 27% found by usual methods and more than 93% with adaptation.
- Slow change in text of persons: 85% with adaptation and 73% with usual methods.
What is not known
The researchers say that the method does not always work. As new models copy persons, it will not be easy to find text that is known to be from persons. Researcher Manish Raghavan says it is not clear that the problem has a solution.
Insight: A detector for AI text must change all the time
Users make new humanizer tools after each release, and the maker trains the detector again on the new text. A related report from 9 October 2026 tests a detector for AI text watermarks, a different protection against the same problem. It reads as a game with no end: each side must change after the other side changes. A detector that does not change finds less with time.
Sources
Posted