Anthropic has introduced invisible, machine-readable watermarks for text generated by newer Claude models as EU AI transparency rules take effect.
Anthropic has begun embedding invisible watermarks into text generated by its newer Claude models, creating a hidden machine-readable signature that can remain with the content even after it is copied and pasted elsewhere.
The move comes as transparency obligations under the European Union’s AI Act take effect. Article 50 requires AI-generated or manipulated content to be marked in ways that make its artificial origin detectable, with the rules applying from Aug. 2, 2026.
Anthropic says the watermark is woven directly into generated text rather than being added as visible metadata. Users will not see a difference in the wording, readability or appearance of Claude responses. The company says the marker can survive copy-and-paste and may remain detectable after some editing.
The system operates at the model level, meaning the marking is not limited to a particular Claude interface. It is intended to cover supported Claude products and services wherever models with the capability are used.
Anthropic is also using a different provenance approach for supported generated files. File outputs can carry digitally signed provenance information based on the C2PA standard, an open framework designed to help establish the origin and history of digital content.
However, the change does not mean every older Claude model is automatically covered. Anthropic has said models launched on or after Aug. 2 support marking at launch, while the company is working to add the capability to models released earlier. The EU has also provided a transition period for AI systems placed on the market before the Aug. 2 deadline, extending certain compliance requirements until Dec. 2.
The development could have major implications for journalists, publishers, students, marketers and businesses that use AI-generated material. A detectable signature could make it easier for platforms and other organizations to establish whether content originated from an AI system.
At the same time, important questions remain about the technology’s effectiveness. Anthropic has not publicly detailed exactly how much editing is required to remove or disrupt the watermark, leaving uncertainty over how reliable detection will be after extensive rewriting.
About The Author
