Claude AI watermarks will tag new Anthropic model output worldwide

Anthropic says Claude AI watermarks will be built into output from supported models launched in the EU on or after August 2, 2026. The plan covers invisible marks in generated text and signed provenance metadata in supported files, with Anthropic saying the system will operate worldwide across its products rather than only in Europe.

The company is implementing commitments it made under the European Union AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. This is not an immediate retrofit of every Claude deployment: Anthropic says support for models released before August 2 remains in progress under a transition period.

Which Claude outputs will have watermarks?

For supported new models, Anthropic says embedded watermarks will cover all generated text. The marks are applied at the model level, so they are intended to appear whether a customer uses Claude through its API, the Claude app, Claude Code, Claude Cowork or Claude Tag.

Text marks will also apply when supported models are accessed through AWS, Google Cloud or Microsoft Foundry, Anthropic says. The company describes the watermark as imperceptible and says it does not alter a response’s meaning, quality or readability. Because it is embedded in the text, it should remain when text is copied and pasted and may persist through some edits.

Files use a different mechanism. Claude will attach digitally signed provenance metadata to supported file types including .svg, .png and .jpg. The metadata follows the C2PA open standard and can indicate that a file was processed by Claude, as well as help establish whether the associated metadata has been tampered with. Availability of signed metadata will vary by platform and feature, including some cloud offerings.

What can a Claude watermark prove?

Anthropic is explicit that a positive detection is limited evidence. A mark means material may have been processed by Claude, not that Claude created every idea, data point or piece of underlying text. A human-authored document can retain a mark after Claude has been used to proofread, translate, summarize or convert it.

The inverse also applies. No detected mark does not establish that content was not generated or processed with AI. Anthropic lists older models, heavily edited or translated text, short excerpts, unsupported features and file types, and metadata removed through conversion, re-saving or screenshots as cases where detection can fail.

Anthropic says it will provide tools and technical documentation so users and third parties can identify the marks, but has not yet published those detection details. For operators building on Claude, the rollout adds a provenance signal to outputs, while leaving them to assess their own transparency requirements under the EU rules.

This story draws on original reporting from SiliconANGLE.