Anthropic Expands Claude Text Watermarking Support with Parameter Limits on Select Models
Frontier AI models · Enterprise AI
Anthropic announced the expansion of text watermarking to older Claude models, including Opus 4.5 and Sonnet 4.5. The invisible watermarking adjusts sampling probabilities without adding billing tokens. However, Anthropic noted that Opus 4.6 and Sonnet 4.6 may fail to reliably generate watermarks under non-default sampling parameters, meaning the absence of a watermark does not guarantee text is human-authored.
Anthropic announced this week the expansion of text watermarking to older Claude model versions, following its compliance measures initiated under the EU AI Act. The technology adjusts text sampling algorithms during generation to embed statistical probability patterns, operating without special characters, extra token billing, or personal data. File outputs continue to be tagged using C2PA Content Credentials. While newer generations include system-level watermarking by default, this update extends coverage to Claude Opus 4.6, 4.5, Sonnet 4.5, and Haiku 4.5, across direct platforms and cloud partners like AWS, Google Cloud, and Microsoft Foundry, with Amazon Bedrock slated for November 10. Anthropic warned that Opus 4.6 and Sonnet 4.6 might not reliably embed watermarks when custom temperature, top_p, or top_k parameters are configured, cautioning detection systems that missing watermarks do not conclusively rule out AI generation.
Key takeaways
- Anthropic expanded text watermarking to earlier models including Opus 4.5, Sonnet 4.5, and Haiku 4.5.
- The mechanism alters word sampling probabilities without adding special characters, extra tokens, or personal data.
- Opus 4.6 and Sonnet 4.6 may fail to reliably generate watermarks when non-default sampling parameters are used.
Source：iThome Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human. Translation checked by AI.