aiexpert
Home / News / Brief
Research · Aug 11, 2026, 08:34 PM · 3 sources

Anthropic watermarks Claude-generated text with invisible markers compliant with EU AI Act Article 50

Anthropic announced it will embed invisible machine-readable watermarks into text and files generated by Claude models starting August 2, 2026, in compliance with the EU AI Act's Article 50 transparency requirements. The watermark is embedded imperceptibly directly into the generated text itself, designed to remain detectable when content is copied and pasted elsewhere and to survive light editing. For files, Anthropic uses the C2PA (Coalition for Content Provenance and Authenticity) standard for signed provenance metadata. The watermarking applies globally, not just in the EU.

Anthropic explicitly cautions that a detected watermark is a signal of Claude processing, not proof of authorship or non-human creation. The watermark may not survive heavy editing, and short passages may provide insufficient material for reliable detection. Additionally, users may ask Claude to proofread human-written work, resulting in marked text that originated from a human. Older Claude models released before August 2 will not immediately carry the marking; Anthropic is working to backport the capability during the EU Act's transition period. Detection tools for users and third parties are in development.

The EU's Article 50 transparency obligations took effect August 2, 2026, under the Code of Practice on Transparency of AI-Generated Content. Non-compliance can trigger fines up to €15 million or 3% of total global annual turnover. Multiple companies—including Google, OpenAI, Meta, Microsoft, and Black Forest Labs—have committed to adhering to the code and using multi-layered marking strategies (metadata, imperceptible watermarking, fingerprinting, logging).

For producers and architects building on Claude APIs, the watermark adds metadata about content lineage and processing. For enterprises handling sensitive documents, the invisible marking may complicate internal workflows where Claude is used for drafting, editing, or proofreading. Developers and students relying on Claude-assisted output should understand that resulting text will carry a machine-detectable signal of Claude involvement, even after copying or light revision.

Sources

Everything this brief rests on
  1. 01 Primary source thenewstack.io
  2. 02 interestingengineering.com interestingengineering.com “Anthropic will add machine-readable watermarks to text generated by new Claude models starting August 2, 2026”
  3. 03 turkiyetoday.com turkiyetoday.com “The transparency requirement forms one piece of a broader regulatory framework in Brussels that governs artificial intelligence systems according to risk level”