Anthropic has clarified how it plans to apply invisible watermarks to text generated by Claude, in a move designed to comply with Europe’s AI transparency rules. The company announced that Claude’s text marking system is “a version of the SynthID-Text approach”, referring to an open-source watermarking technology developed by Google DeepMind.
The technology creates detectable patterns by drawing on wording probabilities, allowing machine-readable marks to be embedded within synthetic text without altering how it reads to the human eye. This makes it possible to identify content produced by the AI model, even where the output appears indistinguishable from writing created by a person.
Meeting the requirements of the EU AI Act
The watermarking feature is being introduced to meet Anthropic’s obligations under the European Union’s AI Act. The legislation requires that synthetic audio, image, video, and text include machine-readable marks so that AI-generated material can be recognised. By adopting the SynthID-Text approach, Anthropic is aligning Claude with these transparency requirements as they come into force.
The rules form part of a wider push to make artificial intelligence outputs more traceable, giving platforms, regulators, and users a means to distinguish machine-produced content from human-authored work. Embedding the marks directly into the wording patterns of generated text is intended to make the identification process reliable and difficult to strip out.
C2PA support for Claude images
Alongside the text watermarking, Anthropic is introducing C2PA support for images processed through Claude. The C2PA standard provides a framework for attaching provenance information to visual content, helping to establish where an image originated and how it has been handled.
Together, the two measures extend coverage across both text and image outputs, addressing the categories of synthetic media that fall within the scope of the AI Act. The SynthID-Text system handles the written material, while C2PA addresses the visual side, giving Claude a broader compliance footing under the new European framework.
The SynthID-Text technology underpinning the text feature was created by Google DeepMind and released as open source, allowing other developers and companies to build on the same watermarking method. Anthropic’s decision to base its own system on this approach ties Claude to an established technique already available to the wider AI community.
Source
Image: theverge.com
