Anthropic is now equipping responses from its AI assistant Claude with an invisible digital watermark. Generated image files also receive signed provenance metadata under the C2PA standard. The step implements the EU transparency obligation under Article 50 of the AI Act, in effect since August 2, 2026 – and, according to the company, applies worldwide, not just to users in the EU.
Watermark survives copying and partially survives editing
The process of how Claude marks AI-generated content is described by Anthropic in its own support center. For text, the model weaves the watermark directly into the generated response without altering its meaning, quality, or readability. Because the watermark is part of the text itself rather than separate metadata, it travels along when the text is copied into other programs and, according to Anthropic, can survive minor rephrasing. In practical terms: a paragraph from a Claude response stays traceable even when pasted unchanged into an email or document.
For image files in SVG, PNG, and JPG formats, Anthropic instead uses the open C2PA standard from the Coalition for Content Provenance and Authenticity. The signed metadata records that a file was generated or edited by Claude. All access paths are covered: the Claude web app, the Claude Platform API, the Claude Code developer tool, the Claude Cowork workspace, and Claude Tag for Slack. Models released since August 2, 2026, support the marking from launch; older models are expected to receive the feature gradually, though Anthropic gives no specific date.
EU Code of Practice makes labeling mandatory
The trigger is Article 50 of the EU AI Act, in effect EU-wide since August 2, 2026, which mandates machine-readable labeling of AI-generated content. A voluntary code of practice tied to this article gives signatory companies a presumption of compliance; besides Anthropic, reports indicate OpenAI, Google, Meta, Microsoft, Black Forest Labs, and video provider Synthesia have also signed. Violations of the transparency obligation can be fined up to 15 million euros or three percent of global annual revenue. The so-called Digital Omnibus pushes numerous high-risk system obligations to 2027 and 2028, but leaves Article 50 untouched – the labeling duty for generated content already applies without restriction.
Anthropic says it applies the marking wherever Claude is offered, including outside the EU and through cloud partners such as AWS, Google Cloud, and Microsoft Foundry. For users in Germany, nothing changes about access: the labeling runs automatically in the background, with no additional fee or separate opt-in.
Detection tool still missing, limits remain open
How reliably the watermark can be proven remains unclear for now. Anthropic says it will give users and third parties a tool to detect the watermarks and metadata in the future, with technical documentation still to follow. Until then, neither newsrooms nor platforms can independently verify whether a text actually came from Claude.
Anthropic itself lists several limitations in its support center: a detected watermark only indicates that Claude may have processed a text – not that the model originally authored it, since users can also submit their own writing for summarizing, translating, or proofreading. Conversely, a missing watermark does not guarantee human origin, for instance because older models don’t yet support the feature. Heavy editing, format conversion, screenshots, or very short passages can further obscure the marking.
What matters now is whether Anthropic delivers the promised detection tool quickly, before competitors like OpenAI or Google follow with their own, possibly incompatible methods. Until outside parties can verify the marking independently, the new labeling’s value remains a matter of trust in providers’ self-disclosure.
