Claude text watermarking
Anthropic says future Claude models will embed a statistical text watermark, adopted with other providers to comply with the EU AI Act marking rule.
Watermark biases word choice without changing output quality, adds no tokens or hidden characters, carries no identifying information and is not specific to Claude. Required of providers serving the EU from 2026-08-02 under the Code of Practice. Fable 5.1 release notes also list C2PA content credentials for generated media.
- Date
- Friday, 14 August 2026
- Lab
- Anthropic
- Kind
- feature
- Access
- closed API
Anthropic says other major developers signed the same Code of Practice and will use their own watermarks; those were not checked here.
Sources
This record was checked against its sources on 6 October 2026. How we check