Anthropic Adds Invisible Watermarks to Claude's AI Creations

Anthropic Integrates Invisible Watermarks into AI Creations of Claude

Anthropic has announced that it will begin watermarking texts and images generated by Claude with machine-readable data, aiming to comply with European regulations on AI transparency. “Generated texts will feature embedded watermarks, and generated files will include digitally signed provenance metadata when supported,” Anthropic states on a new support page for Claude. These changes are invisible to the naked eye but will facilitate the detection of content generated by Claude models by users and online platforms.

These updates represent a future commitment rather than an immediate implementation. The new AI watermarking and transparency obligations under the European AI Act, which came into effect on August 2, include a four-month grace period for compliance of existing AI products launched before this date. Thus, Anthropic clarifies that new Claude models will watermark AI-generated content upon their release, but support for its existing models is still under development.

Machine-readable marks will be applied globally to supported Claude models, including Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. Two different watermarking techniques are used: for images processed by Claude, the C2PA provenance metadata standard—already adopted by Adobe, OpenAI, and Google—will be applied to supported files, but the process for watermarking texts generated by Claude is less detailed.

According to Anthropic, an “imperceptible watermark” is woven directly into the text generated by Claude models without altering the meaning, quality, or readability of the chatbot's response. Anthropic does not name this watermarking system but specifies that these textual watermarks will also be applied when Claude models are accessed via AWS, Google Cloud, or Microsoft Foundry.

“Because the watermark is part of the text, it will travel with the text when copied and pasted elsewhere, and may persist through certain modifications,” Anthropic states on the Claude support page. “The watermarking will be applied at the model level, meaning it will be present regardless of the product or Claude surface from which the text originates.”

Anthropic is also working to enable users and other third parties to detect the watermarks and provenance metadata embedded in content generated by Claude, announcing that it will share details about this detection system in upcoming technical documentation. Several tools are already available to detect C2PA metadata, including Google’s Gemini chatbot, but it is unclear whether these will work with files generated by Claude. I have requested clarification from Anthropic.

This is another step toward clear marking of AI-generated texts and images on online platforms, and a potential victory for those who wish to avoid consuming this type of content. Fanfiction readers have already begun building more rudimentary detection systems to signal when Claude tools have been used in fan works on AO3, but these marking systems could be applied much more broadly—if they work at all.

C2PA data is known to be easily removed, sometimes even accidentally when the media carrying it is uploaded to online platforms, and it is unclear how robust Anthropic's textual watermarking solution is. Even the company itself admits that these marking systems are far from infallible, and any content lacking detectable marks could still originate from generative AI models.