Anthropic has signed the EU AI Act's Article 50 code of practice, so new Claude models will mark their own output from day one: an invisible watermark woven into the text, plus C2PA provenance on generated images, worldwide. The detector isn't ready though. Anthropic says it's still working on letting people check the mark, and its own page grants that a detected mark only means Claude may have touched the text, while a missing mark tells you nothing either. Marking shipped, checking didn't, and checking is what makes the rule worth anything.

support.claude.com/en/a…