Anthropic moves to mark Claude-generated content with invisible watermarks

### Anthropic Introduces Invisible Marking System for Claude AI Content

Anthropic is introducing a new system designed to make content generated by its Claude artificial intelligence models easier to identify. The company plans to add machine-readable markings to newer Claude models and the products that use them. Under the system, AI-generated text will contain an invisible watermark that users cannot normally see. Certain files created by Claude will also include digitally signed information showing their content provenance. Anthropic said the initiative is part of its effort to meet transparency requirements under the European Union’s AI regulatory framework. Claude models launched in the EU on or after August 2 will support the marking system from their initial release. The company is also working to add the technology to older Claude models covered by a transition period under the European rules. The watermark will be embedded directly into generated text at the model level. Anthropic said the marking is designed to remain invisible and should not change the meaning, readability or quality of the content. Because the technology operates at the model level, the markings can apply regardless of where users access supported Claude models. The system is expected to cover products and services including Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag. Anthropic also plans to make the technology available globally wherever supported Claude models operate. The company expects the watermark to remain detectable when text is copied and pasted and after certain forms of editing. However, extensive rewriting, paraphrasing, translation or combining the material with other content could make the signal more difficult to detect. For files, Anthropic will use a different approach based on digitally signed provenance metadata. Supported formats such as SVG, PNG and JPG files may include information based on the Coalition for Content Provenance and Authenticity, or C2PA, standard. The metadata can indicate that a file was processed by Claude and potentially show whether it was modified afterward. Availability of signed provenance information will depend on the file type, platform and technical environment. Anthropic said the text-marking system will also work when supported Claude models are accessed through cloud services such as AWS, Google Cloud and Microsoft Foundry. The company is developing tools that could allow users and third parties to check whether text or files contain Claude markings. Anthropic also plans to publish technical documentation explaining the detection process. However, the company cautioned that detecting a Claude mark should not be considered definitive proof that Anthropic created the underlying content. For example, human-written material could receive a Claude marking after being translated, summarized or proofread using the AI system. Similarly, substantial editing or transformation could remove or weaken the detectable signal. The technology may also have difficulty identifying marks in very short text or content generated by older models that do not yet support the system. File provenance information could disappear through format conversions, screenshots or resaving processes. Anthropic therefore emphasized that the presence of a mark can provide a useful indication of how content was processed, but its absence does not prove that material was entirely created by a human. The move comes as regulators and technology companies increasingly seek practical ways to identify AI-generated material. Anthropic said developers using Claude in their own products will still need to determine which transparency obligations apply to their services under the EU AI Act.