Anthropic Claude watermarks roll out globally on supported new models

Anthropic Claude watermarks will be applied worldwide to output from supported models launched in the EU on or after August 2, 2026, the company says. The move gives Anthropic a machine-readable provenance signal across its products, but it does not yet mean every response from every existing Claude model is marked.

Anthropic says it signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. Models meeting the launch-date threshold will support marking from release, while models released earlier are in a legal transition period and are still being retrofitted. The help-center article does not give a timetable for either that retrofit work or the promised detection documentation.

The policy applies wherever supported Claude models are offered, including the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. Anthropic also says embedded text marks will remain in place when supported models are accessed through AWS, Google Cloud or Microsoft Foundry. Some platforms and product features may not support every type of mark.

How do Anthropic Claude watermarks work?

Anthropic is using separate systems for text and files. For generated text, it says the model embeds an imperceptible watermark in the text itself. The company says the mark does not affect a response’s meaning, quality or readability, travels with copied-and-pasted text, and may survive some editing. The supplied help-center article does not explain the technical implementation of that text watermark.

For supported files, including .svg, .png and .jpg, Claude will attach digitally signed provenance metadata following the Coalition for Content Provenance and Authenticity’s C2PA open standard. Anthropic says a present label indicates that Claude processed the file and can show whether it has been tampered with. File metadata may be unavailable on some platforms, depending on the functions those platforms offer.

A mark is a processing signal, not an authorship verdict

Anthropic’s own limitations are central to how the system should be read. A detected mark means material may have been processed by Claude; it does not establish that Claude originated the ideas, text or data. A person could use the model to proofread, translate, summarize or convert work that began elsewhere, and the resulting output could still carry a mark.

The reverse also holds. An absent detectable mark does not establish that a passage was not generated or processed by Claude. Anthropic lists several cases where detection can fail: output from a model that predates marking, text that has been heavily edited, paraphrased, translated or mixed with other writing, and passages too short to provide a reliable signal. Metadata can also disappear after format conversion, re-saving or screenshots, while some platforms, features and file types do not support a particular marking method.

Anthropic says it plans to support users and third parties in detecting both text watermarks and provenance metadata, with further technical material to come. Developers building on Claude must separately assess what the EU rules require of their own products, the company says. For operators, the practical question is therefore model coverage, output type and how the content was changed after Claude handled it, rather than whether a single detection result can settle provenance.

This story draws on original reporting from The Decoder.