Anthropic is set to integrate covert watermarks into the text produced by its advanced Claude AI models. This innovative feature will embed an undetectable signal within AI-generated content, allowing its origin to be traced even after being transferred from the chatbot.
This initiative follows Anthropic's commitment to the EU AI Act's Article 50(2) Code of Practice concerning Transparency of AI-Generated Content. As detailed in an update on their privacy and legal support documentation, all new Claude models released from August 2, 2026, will incorporate this machine-readable marking from their inception. Although driven by EU regulations, these markings will be universally applied wherever Claude services are available.
The imperceptible watermark is directly embedded into the generated text, designed not to alter the content's meaning, quality, or readability. Since this marking is implemented at the model level, it will function across all Claude platforms, including Claude, the Claude API, Claude Code, Claude Cowork, Claude Tag, and compatible cloud services. Furthermore, the watermark's presence is not contingent on the specific platform where Claude is utilized. Models accessed via AWS, Google Cloud, or Microsoft Foundry will also embed these watermarks, provided the feature is supported.
The watermark is designed to remain with the text even when copied and pasted, and it may endure certain editing processes. However, Anthropic cautions that this technology is not intended as an absolute verification of AI authorship. The detection of a watermark merely suggests that the content might have been processed by Claude, not that Claude was the original author. For instance, AI could be employed for proofreading, translation, summarization, or reformatting human-written material. Conversely, the absence of a watermark does not definitively prove human authorship, as extensive modifications, paraphrasing, translation, brief passages, or combining Claude's output with other texts can render the watermark undetectable. Anthropic has not yet released details on how users or third parties can detect these marks, promising to issue technical guidelines and detection tools in due course.
In addition to text watermarking, Anthropic is incorporating signed provenance metadata into supported file types such as PNG, JPG, and SVG. This metadata adheres to the Coalition for Content Provenance and Authenticity (C2PA) standard, indicating if a file has been processed by Claude. Anthropic joins other major AI developers, including Google and Meta, in adopting the EU's transparency code. This code establishes a framework for adherence to Article 50's mandates regarding the identification and detection of AI-generated content, which became effective on August 2. While existing Claude models launched before August 2 will not immediately support this new marking system, Anthropic is actively working to integrate this feature into these models during the transitional period stipulated by the EU AI Act. This forward-looking approach by Anthropic underscores a commitment to transparency and accountability in the rapidly evolving landscape of artificial intelligence, fostering greater trust and clarity in the origin of digital content. By embedding these subtle yet persistent markers, Claude aims to empower users with tools to better understand and verify the nature of the information they encounter, promoting a more informed digital environment.
