Anthropic has outlined how its Claude artificial intelligence models will mark AI-generated content as it prepares to comply with transparency requirements under the European Union’s Artificial Intelligence Act.

The company said Claude models launched in the European Union on or after August 2, 2026 will support machine-readable marking from launch. Anthropic is also working to introduce the system to models released before that date.

The markings will apply to supported Claude output across its application programming interface (API), Claude, Claude Code, Claude Cowork and Claude Tag, as well as through supported cloud platforms. Anthropic said the system will be used wherever Claude is available, although the type of marking supported can vary by platform and feature.

The system will use two methods: embedded watermarks in generated text and signed provenance metadata attached to supported files.

 

How Anthropic’s Claude watermarking works

For text generated by supported Claude models, Anthropic said an imperceptible watermark will be embedded directly into the output.

The watermark is designed to remain part of the text when it is copied and pasted and may survive some forms of editing. Anthropic said the marking is applied at the model level rather than at the individual product level.

For supported files, including formats such as PNG, JPG and SVG, Claude will attach signed provenance metadata. The metadata will use the Coalition for Content Provenance and Authenticity (C2PA) standard, an open technical standard for recording information about the origin and handling of digital content.

A valid provenance label would indicate that a file was processed by Claude and can also allow users to check whether the associated metadata has been altered.

Anthropic said embedded text watermarks will also apply when supported Claude models are accessed through Amazon Web Services, Google Cloud and Microsoft Foundry. However, signed provenance metadata may not be available on every platform.

 

Also read: Why Nvidia’s Jensen Huang wants Wall Street to finance AI infrastructure

 

What a Claude mark can and cannot prove

Anthropic said it plans to provide tools that allow users and third parties to detect its watermarks and provenance metadata. However, the company said the presence or absence of a mark should not be treated as conclusive proof of whether content was created by artificial intelligence.

A detected Claude mark indicates that content may have been processed by Claude, but does not establish that Claude was the original author.

For example, a person could use Claude to proofread, translate or summarise material originally written by someone else. The resulting text could contain a Claude mark even though the underlying ideas or information came from another source.

Similarly, content can be changed after being processed by Claude. It could be edited, shortened or combined with material from other sources while retaining some form of marking.

The absence of a detectable mark also does not establish that content was not generated or processed by artificial intelligence.

Anthropic said marks could become undetectable if text is heavily edited, paraphrased, translated or combined with other material. Very short passages may also not contain enough text for reliable detection.

For files, provenance metadata can be removed through processes such as format conversion, re-saving or taking screenshots. Marking may also not be available for certain platforms, features or file types.

 

Why the marking matters

The measures form part of Anthropic’s response to Article 50(2) of the European Union Artificial Intelligence Act, which covers transparency obligations for certain artificial intelligence-generated content.

Anthropic said it has signed the European Union’s Code of Practice on Transparency of AI-Generated Content and is using the marking system as part of its commitments under the code.

The company said it will publish further technical information about how Claude’s marks work and how they can be detected.

For businesses using Claude in their own products, Anthropic said they will need to assess separately what transparency obligations apply to their services under Article 50.