Anthropic Rolls Out Claude Text Watermarking to Comply With EU AI Act
Anthropic has announced the global rollout of text watermarking for its Claude models in order to comply with the European Union AI Act. The technology behind this initiative is based on Google DeepMind's SynthID-Text technique, which was published in Nature in 2024.
Artificial intelligence firm Anthropic has officially announced that it will begin implementing text watermarking across its Claude models. This major development is designed specifically to ensure full compliance with the newly enacted European Union AI Act, establishing new regulatory standards for transparency and accountability in artificial intelligence deployment.
The underlying approach utilized for this watermarking system is directly based on Google DeepMind's SynthID-Text technique. This innovative methodology was originally detailed and published in the scientific journal Nature in 2024, providing a foundational framework for identifying machine-generated text securely and reliably.
Technically, the newly implemented watermark relies on a sophisticated mechanism involving a secret key combined with preceding words in the generated output. This specific combination is used to subtly influence word selection during the text generation process, embedding an invisible signature that can later be detected by authorized systems.
Despite the regulatory driver originating within the European Union, Anthropic has confirmed that the watermarking feature will be applied on a global scale. The company explained this worldwide implementation is necessary because it currently lacks a reliable technological method to limit the marking process exclusively to specific geographic regions.
Looking ahead with its deployment strategy, Anthropic has outlined several key expansions for the feature. The company plans to introduce a dedicated detection API and actively bring the advanced watermarking functionality to its older models that were originally released prior to August 2.
