What's this about?
TechCrunch reported that some Claude users are angry that Anthropic's new invisible text watermarks will reveal them using Claude for work and coursework they presented as their own. A technical explainer of how AI text watermarking works circulated the same day, and Ethan Mollick posed the theoretical question: could a superintelligence build a watermarking tool so good that no superintelligence could evade detection? His answer was no.
What exactly is a text watermark?
Rather than marking a file, the watermark is embedded in the word choices themselves. At each step the model has many acceptable next words; the watermarking scheme secretly splits them into preferred and non-preferred groups and nudges the model toward the preferred set. Any individual choice looks natural, but across a long passage the statistical bias is detectable by someone with the key. Because it lives in the words, it survives copy-paste and some editing — but paraphrasing rewrites the choices and generally destroys it.
What's the community saying?
- The complaint is self-incriminating, and everyone noticed. The stated grievance is that a detection feature works — which is a strange thing to publish under your own name.
- Mollick's theoretical point cuts through the practical debate. If no watermark can be made evasion-proof even in principle, then watermarking is a speed bump for casual users, not a guarantee for institutions — while the people most motivated to evade it are exactly those who will.
- The compliance-versus-consumer tension. Anthropic adopted watermarking under an EU AI Act code, but users experience it as surveillance of their own output — the same feature reads as responsibility to a regulator and betrayal to a customer.