Anthropic is introducing machine-readable watermarks for text generated by its artificial intelligence models, including Claude, as the company moves to comply with transparency requirements under the European Union’s AI Act.
The AI company will embed the watermark directly into generated text, making the signal invisible to users while allowing compatible systems to identify content as AI-generated. The measure applies to Claude models released on or after August 2, 2026, the date when the EU AI Act’s transparency obligations for AI-generated content became applicable.
Anthropic is also introducing provenance measures for files generated through its models. For supported files, the company is using C2PA, an open technical standard designed to provide information about the origin and history of digital content.
Unlike visible labels or disclaimers, Anthropic’s text watermark is incorporated into the output itself. This means the watermark can remain with the text when it is copied and pasted to another platform and may continue to be detectable after some editing.
The company said the watermarking mechanism is implemented at the model level. As a result, it is designed to be present regardless of the Claude product or interface used to generate the content. The measure is expected to cover Anthropic's products and services, including Claude, its API and Claude Code.
Anthropic also plans to extend watermarking support to older models. Under the EU framework, certain AI systems placed on the market before August 2 have a transitional period for meeting the marking and detection requirements.
The development follows the implementation of Article 50 of the EU AI Act, which requires providers of generative AI systems that produce synthetic audio, images, video or text to ensure that such content is marked in a machine-readable format and can be detected as artificially generated or manipulated, subject to specified exceptions.
The European Commission has also introduced a Code of Practice on Transparency of AI-Generated Content to help companies comply with these requirements. Adherence to the code is voluntary, but providers covered by Article 50 remain responsible for meeting the applicable legal requirements.
Anthropic is not the only technology company adjusting its systems to meet the European framework. Companies including Google, Meta, Microsoft, OpenAI, Black Forest Labs and Synthesia have also committed to the EU's transparency code.
The wider adoption of watermarking comes as platforms, regulators and AI developers seek technical methods to distinguish synthetic material from human-created content. Google, for instance, has previously developed SynthID technology for identifying AI-generated content across different formats.
Text watermarking, however, has technical limitations. Detection can become more difficult when generated material undergoes substantial rewriting or modification. The absence of a detectable watermark also does not necessarily establish that a piece of content was created entirely by a human.
For Anthropic, the rollout places content provenance directly within Claude's generation process as regulatory requirements around AI transparency become more defined. As generative AI becomes increasingly embedded across content creation, coding and enterprise workflows, machine-readable identification is emerging as one of the mechanisms being adopted to establish the provenance of synthetic content.
Related Articles
