Anthropic Adds Watermarks to 9 Claude Models for EU AI Act Compliance
Updated
Updated · O'Reilly Media · Oct 22
Anthropic Adds Watermarks to 9 Claude Models for EU AI Act Compliance
3 articles · Updated · O'Reilly Media · Oct 22
Summary
Anthropic has embedded AI text watermarks in all outputs from nine current Claude models, covering responses generated through Claude Desktop, Claude Code and its API.
The move is aimed at meeting the EU AI Act, which since Aug. 2 requires providers serving the bloc to mark AI-generated content; Anthropic said other major developers are adopting similar measures.
Claude’s watermark uses Google DeepMind’s open-source SynthID-Text, hiding a detectable signal in word-choice patterns rather than adding visible markers, extra characters or extra tokens.
Anthropic said the watermark indicates whether Claude likely helped write a passage but does not identify the user account; Google testing cited in the report found only a 0.01% change in user approval.
OpenAI and Google are also rolling out watermarking and detection tools, underscoring a broader shift toward labeling AI-generated text, images, audio and video under tightening EU and California rules.
Could the subtle statistical shifts used to secretly watermark AI text unintentionally alter the tone or creativity of your most important documents?
If simple paraphrasing erases Claude's invisible watermark, will this new tracking tool only catch honest users while letting sophisticated scammers slip by?