Anthropic will mark Claude's texts with invisible watermarks in accordance with EU regulations

Watermarking will be added at the model level for all new Claude releases. The system will affect users worldwide, not just those in the European Union. The watermarking method will vary depending on the output type: invisible watermarks are planned for text materials, while supported files will receive a digital signature with origin information. Claude's text responses will be embedded with a discreet signal that doesn't affect readability and remains in the text even when copied and pasted. The developers expect that some of these labels will persist even after editing. This approach will cover all products in the Claude ecosystem, including the API, the chat interface itself, the Claude Code and Cowork tools, and the Claude Tag solution. Enterprise clients working with the model through major cloud platforms—AWS, Google Cloud, and Microsoft Foundry—will also receive labeled output. Anthropic believes this will provide third-party platforms and organizations with an additional technical method to verify whether a particular text fragment was processed by their model. In the future, the company promises to publish technical details and tools for detecting these watermarks, allowing users and verifiers to independently confirm the presence of the markings. A separate mechanism will be used for files, including popular image formats. A digital signature will be embedded in the metadata of generated images or documents, in accordance with the C2PA (Coalition for Content Provenance and Authenticity) standard. This technology creates a single record of content origin: a valid signature confirms that the file was processed by Claude and indicates whether its authorship data has been altered. Anthropic cautions against taking any of these signals as definitive proof of authorship. Claude can process materials originally created by humans. Someone could use the model to translate, shorten, or edit existing text. The resulting material could still bear the Claude label. The converse is also true: the absence of a watermark does not mean the content is human-created. Extensive editing can weaken a text watermark. Short fragments may also prove insufficient for reliable detection. Content created by older versions of Claude will not include the new labeling. Anthropic plans to release additional technical materials as its labeling and detection systems continue to evolve and improve.