Anthropic says Claude text will get invisible watermarks as it prepares for EU AI Act transparency rules.
Why it matters: AI makers and enterprise users will have to account for machine-readable labeling of synthetic text in the EU. The requirement becomes enforceable Aug. 2, 2026, pushing watermarking from a policy idea into product design.
- Anthropic says its Claude text watermarking is based on "a version" of Google DeepMind's SynthID-Text approach.
- The watermark changes token probabilities during generation to create a detectable pattern without visibly changing the text.
- EU transparency rules cited in the story take effect Aug. 2, 2026, and require machine-readable marks on synthetic content.
- The European Commission says the labeling duty does not cover standard editing or content that does not substantially alter the input or semantics.
Anthropic says Claude's text will use an invisible watermarking system as the company prepares for EU AI transparency rules. The company told The Verge that the approach is based on "a version" of Google DeepMind's SynthID-Text system, which works by changing token probabilities during generation so the output carries a detectable pattern.
Google DeepMind has described SynthID as a way to identify AI-generated content while preserving quality and usefulness. Anthropic says its version should not raise user costs or materially affect Claude's output quality.
The policy backdrop is the EU AI Act's transparency regime. The law itself sets Aug. 2, 2026 as the start date for the transparency rules on synthetic content. The European Commission says providers must apply a machine-readable mark to synthetic content and enable detection, while also saying the labeling duty does not cover standard editing or content that does not substantially alter the input or semantics.
Anthropic has already said on its Transparency Hub that it is continuing to explore watermarking and related provenance technologies, but it had not previously described this Claude text mechanism in detail.
The Commission says national market surveillance authorities will enforce the rules in general, while the AI Office oversees systems under its supervision and the European Data Protection Supervisor covers EU institutions. The law also sets fines of up to €15 million or 3% of worldwide annual turnover.
By the numbers
- Aug. 2, 2026 - Start date for the EU AI Act transparency rules on synthetic content.
- €15 million or 3% of worldwide annual turnover - Maximum fines under the law.
What's next: The EU AI Act transparency rules on synthetic content take effect Aug. 2, 2026.