Anthropic Deploys Invisible Watermarks Across Claude Ecosystem for Global Compliance

Anthropic integrates statistical watermarks into all Claude models to meet EU AI Act transparency rules. The embedded signal persists through copy-paste, resisting minor edits but failing against substantial rewrites or translations.

Woofun AI reports that Anthropic has initiated the deployment of machine-readable "invisible watermarks" across its entire Claude ecosystem, including Claude Code, Cowork, API services, and models hosted on AWS, Google Cloud, and Microsoft Foundry. This implementation aims to satisfy the transparency mandates of the EU's AI Act, with plans for global rollout. New models feature this capability natively, while existing versions are being updated incrementally.

The watermark is embedded directly within the generated text rather than as file metadata, ensuring it survives copy-and-paste operations. While the specific algorithm remains undisclosed, official descriptions indicate a mechanism similar to Google SynthID, likely altering token selection probabilities to leave a statistical signature. Anthropic notes that detection signals may be absent in very short texts and can be disrupted by significant rewrites, translations, or content mixing. Consequently, minor word changes are insufficient to remove the watermark, though detection only suggests potential processing by Claude rather than confirming full authorship.