Claude has turned a long-running policy question into a product issue: can AI-generated text be marked in a way machines can read?
QbitAI reported on August 11 that Anthropic is adding invisible marks to Claude-generated content. Anthropic says Claude models launched in the EU on or after August 2 support machine-readable marking from day one. Generated text carries embedded watermarks, while supported files carry signed provenance metadata.
The official wording is concise:
Generated text will carry embedded watermarksIn user terms, Claude output may include a signal that survives copying, pasting and some editing.
Watermarks move into the model layer
The coverage is broad: Claude, Claude Platform API, Claude Code, Claude Cowork and Claude Tag, plus access through AWS, Google Cloud and Microsoft Foundry. Text uses embedded watermarks; supported .svg, .png and .jpg files use C2PA metadata.
The EU rule sets the clock
The legal background is Article 50 of the EU AI Act. CSA breaks it into duties for interactive AI disclosure, machine-readable synthetic-content marking, biometric or emotion-recognition notices, and deepfake or public-interest text disclosure. Transparency duties apply from August 2, with a transition window for older systems.
A detected mark is not proof of authorship
Anthropic also states the limits. A mark only indicates content may have been processed by Claude. It does not prove Claude authored it. Missing marks do not prove the absence of AI either, because older models, heavy editing, translation, mixed text, short passages or stripped file metadata can break detection.
For Chinese users, the practical risk is workflow ambiguity. Teams use Claude Code, meeting-note cleanup, legal polishing and translation. If platforms treat a detected mark as AI authorship, ordinary assisted work may need an appeal path.
Three tests come next
The next checks are public detector documentation, robustness under translation and rewriting, and how platforms explain results. Claude has placed a compliance probe inside text. Whether it becomes useful infrastructure depends on measurement and governance.
Sources: QbitAI, Anthropic Claude Help Center, The Verge, Cloud Security Alliance, CocoLoop, iThome; checked model coverage, product surfaces, cloud access, C2PA file types, Article 50 duties, transition window, detection limits and technical-documentation status.