Anthropic Will Watermark Claude Text to Comply With EU Law
Anthropic will integrate invisible text watermarking into its Claude models to comply with the EU AI Act, enabling downstream verification without sacrificing output quality or speed.
Anthropic announced it will integrate text watermarking into future Claude models, with plans to retroactively update older models launched before August 2, 2026. The move is designed to comply with the European Union AI Act, which mandates that AI providers mark generated content. Having signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026 alongside roughly 190 other signatories, Anthropic is deploying the watermark globally because it cannot yet restrict the feature by region.
The company is using a version of the SynthID-Text framework, which Google DeepMind published in a 2024 Nature paper, building on a 2022 proposal by Scott Aaronson. Instead of using a standard random number generator to select the next word from a list of likely candidates, the system uses a cryptographic key and the preceding words to guide the selection. This leaves a subtle pattern that is imperceptible to human readers but detectable by anyone holding the key. Anthropic plans to release a watermark detection API to help users verify Claude's involvement.
The watermarking technique does not add extra tokens, increase costs, or slow down generation. However, it has limitations. The system cannot easily watermark highly factual statements, code, or minor edits where word choices are strictly constrained, such as completing mathematical equations. It also requires longer passages to achieve high detection confidence. For image and vector files like .png, .jpg, and .svg, Anthropic will instead append C2PA metadata credentials rather than embedding a text-style watermark.
For developers and enterprise users, this implementation preserves the utility of Claude. Internal evaluations and DeepMind's real-world testing on Gemini traffic showed no statistically significant impact on creativity, readability, or user ratings. Furthermore, the watermark contains no identifying information, meaning it cannot trace outputs back to specific users, organizations, or individual chat sessions, thereby preserving user privacy while meeting regulatory demands.
This is our own summary of reporting by Anthropic
