Anthropic detailed its plan on August 10, 2026, to mark AI-generated content from Claude under the European Union’s transparency framework. The company signed the EU AI Act Article 50(2) Code of Practice on Transparency of AI-Generated Content and says new Claude models launched in the EU on or after August 2 include marking support from their release.
The system uses two techniques. Generated text receives an imperceptible watermark embedded at model level, intended to remain with text when it is copied and pasted and to survive some editing. Supported files, including SVG, PNG and JPG formats, can receive digitally signed provenance metadata following the Coalition for Content Provenance and Authenticity, or C2PA, standard. That metadata records processing provenance and can indicate whether a file was altered.
Marks are intended to apply across supported models in the Claude API, consumer Claude service, Claude Code, Claude Cowork and Claude Tag. Text watermarks also apply when those models are reached through AWS, Google Cloud or Microsoft Foundry. Signed file metadata may not work on every partner platform because available file-processing features differ. Anthropic says marking will operate worldwide wherever supported Claude models are offered, not only in the EU.
Models released before August 2 are covered by a transition period. Anthropic is working to add support to those systems but had not provided a completed schedule. The company also plans detection tools for users and third parties, with detailed technical documentation still forthcoming.
A detected mark is a signal, not complete proof of authorship. Claude may have proofread, translated, summarized or reformatted material created by someone else. Marked output can also be edited, excerpted or combined with other content after processing. Detection therefore indicates possible Claude involvement without establishing where every underlying idea or passage originated.
Absence of a mark is equally inconclusive. Older models may not support marking, heavy editing or translation may weaken a text signal, and very short passages may not contain enough information for detection. File metadata can disappear through screenshots, conversion or re-saving, while some products and formats do not support every marking method.
Anthropic advises developers embedding Claude in their own products to assess their Article 50 duties independently. The company’s marking features may help those services meet transparency requirements, but they do not transfer legal responsibility or eliminate the need for contextual assessment. The approach provides interoperable provenance signals while explicitly stopping short of a universal AI-content detector. Its practical reliability will depend on forthcoming detection guidance and on how well marks persist through ordinary editing across different languages and formats.
