Anthropic explains how Claude’s invisible text watermarks will work
| Source: The Verge AI
Tags: Anthropic, Claude, SynthID-Text, EU AI Act, watermarking, AI transparency, C2PA
Anthropic has confirmed Claude will embed invisible text watermarks based on Google DeepMind's SynthID-Text — manipulating low-stakes word choices to leave a key-detectable statistical pattern — to comply with the EU AI Act's mandate that AI-generated text carry machine-readable marks, with no impact on output quality or cost.
Details
Anthropic has detailed how Claude will generate watermarked text under the EU AI Act, which mandates machine-readable marks in AI-generated content. The approach adopts 'a version of the SynthID-Text approach' developed by Google DeepMind, using a key-seeded random process to steer subtle, semantically neutral word choices — rather than arbitrary randomness — creating a statistical pattern invisible to readers but detectable by key holders. The watermarking works at the token-selection level: when multiple plausible completions exist (e.g., 'overcast' vs. 'grey'), the system uses a cryptographic key and preceding context to deterministically choose among equally valid options, leaving a fingerprint without distorting meaning. Anthropic states this adds no cost and has no practical effect on output quality. The same EU obligation applies to other frontier providers. Google's Gemini has supported SynthID since 2024; OpenAI has not disclosed text watermarking plans but is equally subject to the AI Act. Alongside text watermarks, Anthropic is also adding C2PA metadata to Claude-processed images. Practitioners deploying Claude in Europe should note that watermarked outputs enable downstream AI-detection at scale — a capability regulators, media organizations, and enterprises are increasingly demanding.