Research question
Are text detectors and watermarks robust to paraphrasing and spoofing?
Key findings
- Recursive paraphrasing defeats watermarking, zero-shot and trained detectors - Theoretical bound: as models mimic humans, detection approaches random