New transparency regulations for AI-generated content have recently gone into effect in Europe, and Anthropic is responding. Going forward, Claude will embed invisible watermarks. However, this will not only apply when AI-generated content is copied verbatim. On top of that, the AI watermark will remain embedded even when texts are copied, pasted, or partially edited.
Watermarks Instead of Traditional Metadata
Anthropic has decided not to tag its generated content with traditional metadata. After all, this information is usually lost when text is generated and then copied and pasted elsewhere. Instead, Anthropic plans to use a marking system that is embedded directly within the text itself. This type of watermark is not intended to be visible at a glance.
This is the company’s response to the new transparency regulations that will take effect on August 1, 2026, under the EU AI Act. According to these regulations, AI-generated content must be identified as such. Consequently, this watermark will also be used in all Claude models released on or after August 2, 2026. Older models will follow over time.
Furthermore, the marking is not limited to the consumer chatbot: According to Anthropic, it applies wherever supported models are used—such as in Claude Tag, Claude Code, Cowork, via the API, or with cloud partners like AWS, Google Cloud, and Microsoft Foundry. Anthropic has not yet disclosed exactly how the process works from a technical standpoint.
Watermarking Is Not Proof of AI Authorship
However, a key issue remains. A detected watermark does not automatically prove that Claude actually wrote the text. Rather, even the mere use of Claude can trigger a watermark. Anthropic itself points out that texts written by humans can also receive a watermark.
This could be the case, for example, if the text was edited, translated, or reworded by Claude. Conversely, the absence of a mark does not mean that no AI model was involved. Extensive rewriting, for example, can cause this. Anthropic plans to make detection tools available to third parties at a later date.
Images and files follow a different standard
However, as is well known, Claude can be used to generate more than just text. This raises the question of how AI-generated images and other file formats are handled. The text watermark is not used here. Instead, Anthropic plans to use cryptographically signed provenance information based on the C2PA standard.
This technology is already in use across various industries and can be verified independently of the new text-based method. It remains to be seen how helpful Anthropic’s system will be in practice. After all, upon closer inspection, it becomes clear that a mark does not automatically mean the content actually originated from an AI.
In the coming months, the AI leader plans to release additional tools for AI labeling. Anyone interested in the further development of Claude can find the latest background information in our articles on Claude Opus 5 and the free Claude features available without a subscription.
