Copy-paste no more: Anthropic puts invisible watermarks on Claude text under EU rules

Anthropic will add machine-readable watermarks to text generated by new Claude models starting August 2, 2026. The move responds to transparency requirements under Article 50 of the EU AI Act. Anthropic has signed the European Union’s Code of Practice covering AI-generated content. New Claude models launched from that date will mark generated content at the model level. Anthropic says those markings will apply worldwide, not only to users in Europe. The system will use different techniques for text and supported files. Text will receive invisible watermarks, while files can carry signed provenance information. Anthropic plans to embed an imperceptible signal directly into Claude-generated text. Users will not see the mark during normal reading. The watermark will remain part of the text when users copy and paste responses. Anthropic also expects some marks to survive editing. Applying the mark at the model level gives it broader reach across Claude products. Supported output will carry the signal through the API, Claude, Claude Code, Claude Cowork, and Claude Tag. Cloud customers will also receive marked output when they access supported Claude models through major platforms. Anthropic specifically lists AWS, Google Cloud, and Microsoft Foundry. The approach could give platforms and other organizations another way to assess whether Claude processed a piece of text. Anthropic says it will publish technical details for detecting its watermarks. Those tools will help users and third parties check content for supported Claude marks. Older Claude models will not immediately receive the same treatment. Anthropic says it is working to add marking capabilities during the EU AI Act’s transition period. Files get digital provenance Claude will use a separate system for supported files, including common image formats. Generated files can receive signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. C2PA provides a common framework for recording information about digital content. A valid signed record can show that Claude processed a file and indicate whether someone altered its provenance data. That distinction matters because metadata and text watermarks face different technical challenges. A text watermark can travel with copied content. File metadata can disappear during format conversions, screenshots, or re-saving. Anthropic also cautions against treating either signal as definitive proof of authorship. Claude can process material originally created by people. Someone might use the model to translate, summarize, or edit existing work. The resulting text could still carry a Claude mark. The reverse problem also applies. Unmarked content does not necessarily mean a human created it. Heavy editing can weaken a text watermark. Short passages may also provide too little material for reliable detection. Files can lose provenance information through ordinary processing. Older Claude models may also produce content without the newer marking system. For US businesses, developers and content platforms, the change creates another technical signal for tracking AI-generated material. It also gives Claude output a machine-readable identity that can survive across some downstream workflows. Anthropic plans to provide more technical documentation as its marking and detection systems mature.