Anthropic Adds Invisible Watermarks to Claude Content

Core News & Key Facts

Anthropic is introducing machine-readable markings across content generated or processed by supported Claude models, creating an invisible signature that users cannot see with their eyes. For text, the company says Claude will embed a watermark directly into the content. The mark is designed to remain invisible while potentially surviving copy-and-paste operations and some forms of editing without changing the response’s meaning or quality. For files including PNG, JPG and SVG formats, Claude will attach signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, open standard. That metadata can indicate that a file was processed by Claude and can also provide information about whether the file has been altered.

Context & Statements

Anthropic said the marking system is being introduced as the company moves toward compliance with transparency requirements under Article 50 of the European Union’s AI Act. The system will apply to supported Claude models wherever the service is offered worldwide, including Claude Platform, Claude, Claude Code, Claude Cowork and Claude Tag, as well as cloud services involving AWS, Google Cloud and Microsoft Foundry. Anthropic said new Claude models will include content marking from their launch, while models introduced before August 2, 2026 are still being updated. The company also said earlier models have a transition period under the law. Online reaction has been mixed, with some users criticizing the change and others welcoming it as a way to identify undisclosed AI-generated content.

Claude's AI text now carries an invisible watermark

Impact & What’s Next

Anthropic is developing detection mechanisms that will allow users and third parties to check whether supported text or files contain Claude markings. The company said it plans to release more information about those mechanisms soon. However, Anthropic cautions that the system has important limitations. A detected watermark indicates that content may have been processed by Claude, but it does not prove Claude was the original author because users may employ the service for proofreading, translation, summarisation or file conversion. Conversely, the absence of a mark does not prove that content was not created or processed by AI. Older models may lack marking support, while heavy editing, paraphrasing or combining content can remove or obscure a mark. Anthropic’s approach is similar in purpose to Google DeepMind’s SynthID system for AI-generated text.