Supervised Drafting and the AI Detector
The paper makes two points about AI detectors.
- Relying on a detector’s reported accuracy commits the prosecutor’s fallacy. Vendors report how often AI-written text is flagged, but an accuser needs the chance that a flagged text was written by AI. That chance is about 94% for native English writers and only 20% for non-native ones.
- A detector may confuse AI-assisted writing with sole AI authorship. It cannot tell an author who directed and checked an AI’s draft from one who let the AI write alone.