Overview
- Anthropic has rolled the watermark into Claude models launched after the EU rule took effect and says older models will be marked by December, and the change is being applied worldwide because the company cannot limit it by region.
- The watermark works by nudging the model’s token selection to create a subtle, probabilistic pattern that only a machine with the detector key can reliably test for, not a visible tag a reader can see.
- Detection is probabilistic and sample‑size dependent, so the mark shows likelihood of Claude involvement rather than how much the model was used or whether a human originated the content.
- Independent developers have already published methods that weaken or remove the statistical signal, and Anthropic has not yet released its detector, key, or empirical error rates, raising concerns about false positives and centralized verification power.
- The move deepens practical questions for educators, employers, and businesses over acceptable AI edits, IP ownership, and access to detection tools, and it joins similar efforts by Google, OpenAI, and Meta as provenance systems become an industry norm.