Anthropic to Watermark Claude Text and Files Worldwide for EU AI Act

Anthropic will embed watermarks in text generated by future Claude models and attach signed provenance metadata to supported files, extending European Union transparency rules to Claude users worldwide.

The company said the changes are meant to comply with the EU AI Act, whose transparency obligations for certain AI-generated and AI-manipulated content apply from August 2, 2026. Marking will not be limited to the EU. It will apply across Claude products and access points, including Claude, its API, Claude Code, Claude Cowork, and Claude Tag.

Applying the rules at the model level pushes a European standard into global software and content workflows, and adds engineering and governance costs for AI providers.

In a public explainer, Anthropic said "future Claude models will generate text that contains a watermark" and described the system as "a way of determining the likelihood that Claude was involved in writing the text." The company said it, along with other major AI providers, is implementing the change to comply with the EU AI Act.

Anthropic said the text watermark is imperceptible and does not alter meaning, quality, or readability. It carries no identifying information and cannot be traced to a specific person, organization, or chat, the company said, and the watermarking method will not be unique to Claude.

Because the watermark is part of the text itself, it can persist when text is copied and pasted and may survive some editing. New models will mark AI-generated content from launch, according to Anthropic's help documentation. For models launched before August 2, 2026, the company said EU law allows a transition period and that it is working to add watermarking to those systems.

For files, supported generated outputs will include signed provenance metadata. Verified materials point to formats including .svg, .png, and .jpg.

The EU rules reach beyond model providers. European Commission materials say the transparency obligations cover both providers and deployers and require certain AI-generated or manipulated content to be clearly labeled. That includes deepfakes and AI-generated or AI-manipulated text published on matters of public interest. The Commission has also published icons for use in labeling AI-generated content.

For enterprise users, the effect may reach beyond disclosure notices. Watermarking and provenance metadata can affect content review pipelines, audit processes, and downstream detection systems used by customers and distribution partners, including cloud channels where Claude is offered such as Google Cloud, Microsoft Foundry, and AWS.

The announcement points to two broader themes. One is rising compliance spend as frontier AI developers build controls into models and delivery systems. The other is fragmentation risk, as rules written in one jurisdiction set a de facto baseline for global deployment, procurement, and content standards.