Anthropic has begun embedding invisible watermarks into text produced by its newer Claude models, applying the measure globally as the company prepares to comply with the European Union's AI Act, which requires providers of AI systems that generate synthetic content to ensure their outputs are detectable.

Claude models released from 2 August onward will carry an imperceptible signal at the model level. According to Anthropic, the watermark does not alter the meaning or readability of the text and remains embedded when content is copied and pasted. The company noted that the mark "may persist through some editing." Anthropic has signed the EU's Code of Practice on AI-generated content and stated that the watermarking will apply worldwide rather than only to users within the EU.

How the Watermarking System Works

The signal will be applied across multiple surfaces, including the Claude app, the API, Claude Code, Claude Cowork, and Claude Tag. Customers accessing supported models through Amazon Web Services (AWS), Google Cloud, and Microsoft Foundry will also receive watermarked output.

File outputs are handled through a different mechanism. Supported file formats, including common image types, will carry signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) standard, rather than an embedded text watermark. C2PA is a cross-industry specification developed jointly by Adobe, Microsoft, the BBC, and others, and is increasingly being adopted as a common framework for verifying the origin and editing history of digital media.

Older Claude models will not immediately receive the watermarking system. Anthropic said it is working to extend the capability to those models during the EU AI Act's transition period.

The company cautioned that heavy editing, paraphrasing, translation, or blending Claude's output with other text could render the watermark undetectable. It also noted that detecting a watermark does not definitively confirm that Claude originally authored a given piece of text, since activities such as proofreading or translating existing work through Claude can also leave a mark.

Anthropic plans to publish technical details to assist users and third parties in detecting its watermarks, with further documentation expected as both the marking and detection systems continue to evolve.

Anthropic is the second major AI lab to introduce text watermarking. Google DeepMind announced in 2024 that it was watermarking text and video generated through the Gemini app and web experience using its SynthID technology, having previously released an image-watermarking version of the tool in 2023. Other major providers, including OpenAI and Meta, have not yet deployed comparable text-level watermarking, though the EU AI Act's transparency requirements will apply broadly to all general-purpose AI model providers operating in the region.