Anthropic has introduced invisible, machine-readable watermarks for text generated by its Claude AI models, enabling users and third parties to identify content that has been processed by Claude even after it is copied and pasted across platforms.
The company disclosed the development in an updated Claude Help Center article, noting that Claude models launched on or after August 2, 2026 will support the watermarking system from launch.
The watermark is embedded directly into generated text and remains invisible to readers. According to Anthropic, it is designed not to affect the meaning, quality or readability of the content.
Watermark designed to travel with text
Unlike watermarks associated with a specific application or interface, Anthropic's system operates at the model level. As a result, the signal can travel with text when it is copied to other platforms and may remain detectable through certain forms of editing.
The watermarking system will be available across supported Claude models and Anthropic's products and services, including Claude, its API, Claude Code, Claude Cowork and Claude Tag. It will also apply when Claude is accessed through third-party platforms such as AWS, Google Cloud and Microsoft Foundry.
Anthropic said the initiative forms part of its commitments under Article 50(2) of the European Union's AI Act Code of Practice on Transparency of AI-Generated Content.
Provenance metadata for AI-processed files
Alongside text watermarking, Anthropic is introducing a separate provenance mechanism for certain files generated or processed through Claude.
Supported formats, including SVG, PNG and JPG files, will carry digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity (C2PA) standard. The metadata can indicate whether a file was processed by Claude and help establish whether its provenance information has been modified.
However, Anthropic noted that signed provenance metadata may not be available when Claude is accessed through every platform.
Watermark is not an AI authorship detector
Anthropic has cautioned that the watermark should not be interpreted as definitive proof that content was authored by AI.
A detected Claude watermark only indicates that the content may have been processed by Claude. For example, a user could submit their own writing to Claude for proofreading, translation or summarisation, after which the resulting text could carry the watermark despite the underlying ideas originating with the user.
Conversely, the absence of a watermark does not necessarily establish that content was written without AI assistance. Anthropic said the signal may become undetectable after heavy editing, paraphrasing, translation or when text is combined with other material. Very short passages may also not contain enough text for reliable detection.
The company therefore describes the technology as a provenance signal rather than a standalone AI-content detection system.
Anthropic said it is developing tools that will enable users and third parties to detect the embedded watermarks in Claude-generated text, alongside tools for identifying provenance metadata attached to supported files.
