Anthropic Reveals How Watermarks Work in Claude
Photo: TechCrunch
Quick answer
Anthropic introduces watermarking in Claude’s AI-generated text to comply with the EU AI Act, ensuring traceability without compromising content quality.
Anthropic has shared details about its watermarking technology for text generated by its AI model, Claude. This initiative aligns with the European Union’s requirements for transparency in AI-generated content. The watermarks remain invisible to users but enable origin identification through a specialized algorithm.
The technology is based on Google DeepMind’s SynthID-Text approach. Unlike methods that analyze stylistic features, watermarks are embedded via minor changes in word or phrase selection. Anthropic plans to release an API for detecting these watermarks, simplifying content verification.
User reactions to the innovation vary. Some see it as essential for combating misinformation, while others view it as an unnecessary restriction. The company emphasized that light editing won’t fully remove watermarks, though a complete rewrite may render them ineffective. In code generation, watermarks are applied minimally—primarily in comments—to avoid disrupting functionality.
Anthropic noted that other major AI developers are also adopting similar technologies to meet transparency obligations. This suggests watermarks will become a standard for identifying AI-generated content.
Common questions
- How do watermarks work in Claude?
- Watermarks are embedded through subtle, undetectable changes in word or phrase selection, detectable only via a specialized key. This method does not affect content quality.
- Can watermarks be removed from text?
- Minor edits won’t fully eliminate watermarks, but a complete rewrite with word replacements may render them ineffective.
- How do watermarks impact code generation?
- Watermarks are minimally applied in code—primarily in comments—to avoid disrupting functionality, as the model must produce executable code.
Dzen feed: /feed/dzen.xml · RSS: /feed.xml