A new essay argues that watermarks baked into AI-generated text will never work because paraphrasing wipes them out instantly.
The post, published this week by developer Sean Goedecke, lays out why text watermarking is fundamentally weaker than watermarking images or audio. Image watermarks survive compression and cropping because they're embedded in pixel data that isn't easily rewritten piece by piece. Text watermarks, by contrast, work by nudging a model's word choices into statistically detectable patterns, a signal that any paraphrase, translation, or light edit can scramble beyond recovery. The essay picked up real traction on Hacker News, pulling in 75 points and 63 comments.
The argument lands as schools, publishers, and platforms keep hunting for a technical fix to tell human writing from AI writing. If a watermark can't survive a five-second pass through a paraphrasing tool, any policy or product built on that promise is offering more certainty than the technology can actually deliver.
It's a reminder that AI text detection is still mostly an arms race, not a solved problem, and this may be one race the defenders keep losing.