Claude now embeds invisible watermarks in AI-generated text
Anthropic is introducing machine-readable marks in Claude to identify AI-generated content in compliance with the EU AI Act. New models from August 2026 will include invisible watermarks in text and signed metadata for files, with marks working across all Claude platforms worldwide. Detection mechanisms will be shared in upcoming technical documentation.
What is the purpose of watermarks in AI-generated text?
Watermarks serve to provide transparency and enable detection that content originates from an AI model. They help meet legal obligations under the EU AI Act and give readers context about the source of information they consume.
Where do the watermarks appear and persist?
Invisible watermarks are embedded directly in the text itself, so they travel when content is copied and pasted elsewhere and may survive certain edits. They remain present regardless of which Claude platform you use.
What are the limitations of watermark detection?
Detecting a watermark is not fully conclusive—it signals that content may have been processed by Claude but does not confirm complete provenance. Additionally, missing watermarks do not necessarily mean content wasn't AI-generated, such as when it has been heavily edited or comes from an older model.
When will watermarking be available for older Claude models?
The law includes a transition period for Anthropic models released before August 2, 2026. Anthropic is working to add watermarking support to these older models, though no specific timeline has been announced.
- Claude Cowork upgrade in Chrome — claude.com 77 % match
- Diagram Design - 29 editorial diagram types for Claude Code — github.com 71 % match
- Anthropic introduces Claude Fable 5.1 and Claude Mythos 5.1 — anthropic.com 69 % match
- Anthropic16
- Claude8
- EU AI Act
- Article 50(2)
- C2PA
- Claude Platform
- Claude Code8