Claude will apply invisible watermarks to AI text and images
Anthropic has outlined plans to apply machine-readable watermarks to content generated by its Claude models, covering both text and images. According to a new Claude support page, generated text will carry embedded watermarks while generated image files will include digitally signed provenance metadata where the format supports it. Neither type of marking is visible to the human eye, but both are designed to be detectable by automated systems and online platforms.
The move is primarily driven by compliance with European Union rules on AI transparency, which are introducing new labeling and disclosure obligations for AI-generated content. By building provenance signals directly into outputs, Anthropic is positioning Claude to meet these requirements ahead of enforcement deadlines. The broader regulatory landscape in Europe - particularly under the EU AI Act - has been pushing AI developers toward more traceable and auditable outputs.
On the image side, the approach aligns with the C2PA standard (Coalition for Content Provenance and Authenticity), an industry-backed specification for embedding signed metadata into media files. C2PA credentials can record information about how a file was created or modified and by what tool, and a growing number of platforms and devices have begun supporting the standard. Text watermarking is a somewhat less mature area technically, but several research and commercial approaches now exist for encoding detectable signals into the statistical patterns of generated language.
It is worth noting that these are forward-looking commitments rather than features already available to users. Anthropic has not specified a precise rollout timeline, and the effectiveness of text watermarking in particular remains an active area of research - determined adversaries can sometimes remove or obscure such signals. Still, formalizing the commitment and publishing it publicly adds a degree of accountability, and the adoption of C2PA for images puts Anthropic in line with steps already taken by other major AI image providers.
