Can AI-Generated Text be Reliably Detected?
Argues that recursive paraphrasing defeats both classifiers and watermarks, and that detection gets harder as models get closer to human text.
What it is
The pessimistic paper. The authors show that recursive paraphrasing, running a rewriter more than once, defeats both classifiers and watermarks of the day, and they argue from first principles that as model text approaches the human distribution, any detector must trade false positives against misses. They also demonstrate spoofing: making human text look watermarked. It is the paper detector sceptics cite, and it is why providers now describe their marks as useful signals rather than proof.
When to use this entry
Cite it for the theoretical limits of detection and for recursive paraphrase attacks.
Questions
Is reliable AI text detection impossible?
This paper argues it becomes impossible as models approach human text; others argue length and watermark keys keep it useful. The honest answer is that detection is probabilistic.