Anthropic says Claude models in the European Union will now add invisible watermarks to AI-generated text and C2PA provenance metadata to generated files, according to Techmeme’s summary of Thomas Claburn’s report in The Register. The stated driver is compliance with the EU AI Act, with the rules cited as a reason to help trace the ancestry of AI-generated output. A monitor-tier item says the policy applies to new Claude models launched in the EU on or after August 2, 2026. That item also says the C2PA metadata applies to supported files, while the text watermarking applies to generated text. The change is best read as a provenance layer, not a guarantee of permanent attribution. The monitor-tier item says the marks are meant to support transparency, but can be lost through editing, screenshots, or file-format conversion. That limitation matters for operators. C2PA metadata can help preserve provenance when files remain in supported formats and workflows, but it is not the same as an indelible marker across every downstream copy. Invisible text watermarking similarly depends on the generated text remaining sufficiently intact. For Anthropic, the move gives Claude a compliance posture in Europe as AI Act obligations begin to shape product behavior. For customers, it means AI-generated output from Claude in the EU may carry provenance signals by default, while still requiring caution when content is transformed, copied, or republished outside the original workflow. Who benefits: EU users and organizations that need clearer provenance signals around Claude-generated output benefit most. Anthropic also benefits by aligning its product behavior with cited EU AI Act requirements. Who's exposed: Teams relying on watermarks as a complete detection system are exposed, because the provided summaries say marks can be lost through editing, screenshots, or format conversion. Workflows that strip metadata may also reduce the usefulness of C2PA provenance.