Anthropic is implementing machine-readable tagging technology for its Claude series of AI models. All text generated by Claude will be marked with an invisible watermark, while images and other files will receive digitally signed provenance metadata. This move signals that regulation around AI content transparency is entering a phase of practical enforcement.
According to an official statement from Anthropic published on August 11, the watermark is woven directly into the text content itself, rather than being attached at the metadata level. This means the watermark continues to propagate when the text is copied and pasted to other platforms, and may persist even after partial editing. This feature significantly enhances the ability to trace AI-generated content, surpassing previous methods that relied on metadata labels.
The immediate trigger for this plan is Anthropic's signing of the EU's Code of Practice on Transparency of AI-generated Content under Article 50(2) of the AI Act. Under this code, new Claude models launched in the EU from August 2, 2026, must support content tagging from the date of release. Anthropic also stated that the tagging mechanism will apply globally, covering all Claude product lines and partner cloud platforms, not just the EU region.
How the watermark works: embedded in text, travels with content
Anthropic uses two complementary technologies to tag content generated by Claude. The first is an embedded text watermark. When a supported Claude model generates text, the model weaves an invisible watermark directly into the text itself. This watermark does not alter the meaning, quality, or readability of the content, and users cannot detect it with the naked eye. Because the watermark is a component of the text, rather than attached metadata, it moves with the text when copied and pasted, and may remain after partial editing. The watermark is applied at the model level, meaning that any text generated through any Claude product or interface will carry this tag.
The second is signed provenance metadata. When Claude generates supported file types, such as .svg, .png, or .jpg, the output will be appended with a digitally signed provenance metadata label. This metadata follows the open standard of the Coalition for Content Provenance and Authenticity (C2PA), which is already widely adopted in the industry for recording content source information. If the signed metadata tag is present, it indicates that the file was processed by Claude and can be used to detect if the file has been tampered with.
Anthropic clarified that the tagging mechanism will cover all major Claude products and access channels, including the Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag. It will also apply to cloud partners accessed through AWS, Google Cloud, and Microsoft Foundry. Geographically, the tagging will apply to all regions where Claude offers services, not just the EU. Regarding timing, new models launched in the EU after August 2, 2026, must support tagging from their release date. For existing models already released, EU regulations provide a transition period. Anthropic stated it is working to add tagging support to these models and will update relevant documentation accordingly.
Technical limitations: a signal, not definitive proof
Anthropic explicitly disclosed the limitations of the watermark technology in its official statement, which is particularly important for users and developers who rely on this mechanism for content verification. Detecting a Claude tag only indicates that the content was "likely" processed by Claude, but it does not fully confirm the complete origin of the content. For example, Claude may not be the original author, as users often use Claude for proofreading, translation, summarization, or file conversion. Even if the output carries a Claude tag, the underlying ideas, text, or data may come from other sources. Additionally, the content may have been modified, excerpted, or merged with other materials after being processed by Claude.
Conversely, the absence of a detected tag does not prove that the content was not AI-generated. The following scenarios can all lead to a missing tag: content generated by a model released before the tagging feature was launched; text that has been heavily edited, rewritten, translated, or mixed with other content; text passages that are too short to carry a reliable signal; file metadata that has been stripped during format conversion, re-saving, or screenshot operations; or content from platforms, features, or file formats that do not yet support the specific tagging type.
Furthermore, Anthropic specifically noted in its statement that developers deploying Claude in their own products should independently assess the requirements of Article 50 of the AI Act for their specific products and services. Anthropic stated that its goal is to support developers in fulfilling their own transparency obligations, and it will release detailed guidance on tagging and detection methods once technical documentation is ready. This means that for enterprises integrating Claude capabilities into downstream products, compliance responsibility for AI content tagging does not automatically transfer due to Anthropic's technical implementation. Developers must still proactively clarify their own obligations under the relevant regulatory framework.