Anthropic has announced that all new Claude AI models released on or after August 2, 2026, will embed an invisible watermark within the text they generate. This initiative aligns with transparency requirements outlined in the EU AI Act’s Article 50 and applies globally across all Claude platforms.
The watermarking system operates at a statistical level, subtly influencing word choice patterns during text generation without altering readability. Unlike visible labels, this hidden signal can be detected only through specialised tools designed to recognise the pattern.
Scope and Functionality
- The watermark covers various Claude services, including the API, chat app, Claude Code, Claude Cowork, and Claude Tag.
- It applies regardless of user location and extends to deployments on AWS, Google Cloud, and Microsoft Foundry.
- The watermark is embedded in the sequence of words, making it resilient to copying, pasting, and minor edits.
Limitations and Considerations
- No public algorithm or detection tool has been released yet, leaving practical effectiveness unverified.
- The watermark confirms that Claude processed the text but does not identify the original author or extent of AI involvement.
- Heavy rewriting, mixing human and AI content, or short text segments may prevent detection.
- Older Claude models predating August 2026 will receive watermarking during a future rollout without a set timeline.
Digital File Provenance
For image formats like PNG, JPG, and SVG, Anthropic will add provenance metadata following the C2PA standard. This metadata tracks file creation and edits but is stored separately from the content and can be removed by common file operations such as screenshots or format conversions.
Anthropic emphasises that these measures aim to enhance transparency rather than establish definitive authorship, urging cautious interpretation of watermark detection as an indicator of AI processing rather than proof of origin.
