How Claude marks AI-generated content

Anthropic has agreed to the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content for both generative AI models and AI systems. This article explains how Anthropic plans to implement these commitments, how marking functions, and its limitations. Updates and technical guidance will be provided as they become available.

Commitments under the Code apply to Claude models launched in the EU on or after August 2, 2026, which will support machine-readable marking from the start. Existing Claude models will also receive marking support. Generated text will include embedded watermarks, and generated files will contain digitally signed provenance metadata when supported.

These markings will be present across various Claude products and services worldwide. The two main techniques used by Claude to mark content are (1) watermarks embedded in the text and (2) signed provenance metadata attached to files. Watermarks are imperceptible and do not alter the content's meaning, quality, or readability. Signatures are only attached to supported file types.

Detection mechanisms will be available to help identify Claude's marks in text or files, though they are not definitive proof of the content's provenance. Some limitations exist, such as the mark not confirming the original author, changes made after processing, or scenarios where a mark may not be present. Organizations using Claude should independently assess their own transparency obligations under the EU Code.