By 
Anthropic has introduced an invisible watermarking system for Claude that is designed to help identify content generated by its artificial intelligence models as governments move towards greater transparency over AI-generated material.
The watermark is embedded directly into text produced by supported Claude models and cannot be seen by readers. Anthropic says the marking is designed to preserve the meaning, quality and readability of the generated text while creating a machine-readable signal that can be used to determine whether Claude was involved in producing it.
Unlike a conventional watermark that appears visibly on a document or image, Claude’s text watermark is incorporated into the statistical patterns of the language generated by the model. Anthropic says the mark can remain with the text when it is copied and pasted and may survive some forms of editing.
The move is linked to the European Union’s Artificial Intelligence Act and its transparency requirements for AI-generated content. Anthropic said it is applying the marking globally because it does not currently have a durable way to restrict the system only to particular regions.
Under Anthropic’s approach, Claude models launched from August 2, 2026 are required to support machine-readable marking at launch, while the company is working to add the technology to older models during the transition period provided under the EU rules.
The system also extends beyond text. For supported generated files, including certain image formats, Anthropic plans to use digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) standard. This creates information about the content’s provenance and can help identify whether the file has been altered.
Anthropic is also developing tools that will allow users and third parties to detect Claude’s marks. The company has announced plans for a watermark-detection API, potentially allowing developers to incorporate Claude-content detection into other applications.
The change could make questions about the origin of digital content increasingly important. A piece of text copied from Claude may look identical to ordinary writing, while still carrying a machine-readable signal that can potentially be detected later.
The technology could also become significant as schools, workplaces, publishers and online platforms develop their own policies on AI-assisted writing. Watermarking provides a potential way to establish whether AI was involved, although it does not by itself determine how much of a person’s work was generated by AI or whether the use of AI was appropriate.
The effectiveness of such systems will ultimately depend on how accurately they can identify AI-generated content, how well the marks survive legitimate editing and how organization interpret the results.
Comments