Anthropic adds invisible watermarks to Claude to meet EU transparency rules
The AI lab is embedding machine-readable signals at the token layer to identify synthetic text even after it is copied and pasted.
Anthropic has introduced invisible, machine-readable watermarks for text generated by its Claude AI models to ensure synthetic content can be identified. The move establishes a technical layer of accountability as generative AI becomes more pervasive across digital communications.
The new system embeds imperceptible signals directly into the text at the token generation layer. According to Anthropic, these watermarks do not alter the meaning or readability of the text for human users, but they allow the content to be identified as AI-generated even after it has been copied and pasted. The company noted that these marks may also persist through some editing.
Regulatory Drivers
This implementation is a direct response to Article 50(2) of the European Union's AI Act, specifically the Code of Practice on Transparency of AI-Generated Content. The EU mandate requires that AI-generated content be clearly identifiable to prevent the spread of misinformation and maintain transparency for end users.
Anthropic is not alone in this approach. The company joins other major AI laboratories, such as Google DeepMind—which utilizes its SynthID technology—in deploying technical safeguards to mark synthetic text. This collective shift reflects a broader industry trend toward systemic accountability and compliance with emerging international legal frameworks.
Industry Implications
By making watermarks invisible yet persistent, AI labs are attempting to close the "detection gap." Traditional AI detectors often rely on statistical patterns that can be easily bypassed or result in false positives. Embedding signals at the generation layer provides a more robust method of provenance that survives basic manipulation.
However, the move raises ongoing questions regarding the permanence of these marks. While they survive copying, the industry continues to grapple with whether sophisticated users can employ specialized tools to strip these signals from the text entirely, potentially undermining the transparency goals of the EU AI Act.
Deployment Timeline
Watermarking is supported at launch for new Claude models released on or after August 2, 2026. To ensure comprehensive coverage across its ecosystem, Anthropic stated it is currently working to add marking support to Claude models that were released prior to that date.