Summary
Anthropic published the technical mechanism behind Claude's invisible, machine-readable text watermarking, applying to all Claude models launched on/after 2026-08-02 (older models in transition).
Why it matters
Developers building detection or provenance pipelines — or redistributing Claude output — need to understand the watermark's semantics. More broadly, watermarking is becoming a standard property of API models.
Technical details
| Applies To | Claude models launched on/after 2026-08-02; pre-existing models in transition |
|---|---|
| Mechanism | embedded in generated text, machine-readable |
Updates
2026-09-02 rollout update (via the 9/1 Fable 5.1 launch): EU AI Act Article 50(2) obligations apply to models placed on the market after 2026-08-02 — Claude Fable 5.1 is the first tracked Anthropic launch shipping with output watermarking enabled, and Anthropic's watermark detection API is now in private preview (access on request). Watermark detection for provenance marking is becoming an integration surface for EU-facing deployments
Tags
anthropicwatermarkprovenancedetection