Claude will mark its content with invisible watermarks and signed metadata — worldwide

Anthropic signed the EU AI Act Article 50(2) Code of Practice on transparency of AI-generated content, and announced how it will start marking everything Claude produces so it can be identified as AI-generated.

It applies worldwide, not just in Europe. While the commitment stems from European regulation, Anthropic clarifies that marks will apply "wherever Claude is offered, worldwide". Because marking happens at the model level, anyone using Claude — in Colombia, the US, the EU or anywhere else — will generate content carrying these marks.

What gets marked and when: Claude models launched on or after August 2, 2026 will include marks from day one. Older models will get support during the transition period. Marks apply across all products (Claude, Claude Code, Claude Cowork, Claude Tag and the API), including when accessed through AWS, Google Cloud or Microsoft Foundry.

Two complementary techniques:

1. Text watermarks: An imperceptible pattern woven into the text itself. It doesn't change the meaning or quality of the response, and it survives copy-paste and some editing.

2. Signed file metadata: For files like .svg, .png or .jpg, Claude attaches metadata following the C2PA (Coalition for Content Provenance and Authenticity) standard — the same one used across the industry to track content origin and detect tampering.

Important limitations: Detecting a mark tells you the content was processed by Claude, but doesn't confirm authorship — many people use Claude to proofread, translate or summarize other people's text. And the absence of a mark doesn't mean it isn't AI: very short passages, heavily edited text, or files converted between formats may lose the mark.

For those building with Claude: Anthropic will publish technical documentation so third parties can detect the marks and meet their own transparency obligations.