Claude invisible watermarks will mark supported new model outputs worldwide

Anthropic says Claude invisible watermarks will be included in generated text from supported new models, with signed provenance metadata added to supported generated files. The policy is tied to Anthropic’s commitment under the EU AI Act’s Article 50(2) transparency code, and it extends worldwide wherever the covered models are offered.

The practical limitation is that the marks cannot yet be independently assessed. Anthropic says it will provide users and third parties with ways to detect both text watermarks and file metadata, but technical documentation on that detection is still forthcoming.

Which Claude outputs will carry invisible watermarks?

Anthropic says Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. It is also working to add marking to models released before that date during the transition period.

For supported models, embedded watermarks apply to generated text across the Claude API, Claude’s consumer product, Claude Code, Claude Cowork and Claude Tag. The company says the same applies when covered models are accessed through AWS, Google Cloud or Microsoft Foundry. Coverage is not uniform across every surface: some platforms and features may not support every kind of mark.

Text marks are designed to be imperceptible. Anthropic says they do not change the meaning, quality or readability of a response, and that they travel when text is copied and pasted, potentially surviving some editing. For supported file types, including .svg, .png and .jpg, Claude will attach signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, open standard.

What does a detected Claude mark prove?

Less than many users may assume. Anthropic says a detected mark indicates that material may have been processed by Claude, rather than conclusively establishing the full origin of the work or proving that Claude authored it. That distinction applies when someone uses the model to proofread, translate, summarize or convert material that originated elsewhere.

The inverse is also true. No detected mark does not show that a document was not generated or processed by AI. Anthropic says watermarks may become undetectable after substantial editing, paraphrasing, translation or mixing with other text, and very short passages may lack enough material for a reliable signal. File metadata can disappear after format conversion, re-saving or a screenshot.

Anthropic describes two marking methods, watermarks for text and provenance metadata for files, as tools for signaling both generated and processed content. Its documentation is clearer on the marking of generated outputs than on every possible treatment of user-supplied material, so the scope of processing-related marks will need technical scrutiny once detection details are available.

The announcement puts Anthropic among AI providers building provenance signals into their products, but it does not create a reliable authorship test by itself. Until detection mechanisms are published and tested, operators cannot measure the marks’ coverage, resilience or false-negative rate from Anthropic’s announcement alone.

This story draws on original reporting from Ars Technica Policy.