Bottom line: Anthropic has disclosed how the Claude watermark works, which is intended to serve the labeling obligation for AI content under the EU AI Act and, according to the company, is difficult to remove.
Anthropic has disclosed how the watermark embedded in Claude-generated texts works and how robust it is against removal attempts. The background is the labeling obligations for AI-generated content stipulated by the EU AI Act.
Anthropic has published details on how the watermark embedded in texts generated by Claude works. According to Golem.de, the company explains where the method used originates from and how difficult it is to remove the marking afterward. The source does not specify concrete technical parameters, such as the detection rate or the underlying algorithm, in detail, but does point out that the technique was specifically designed for high resistance to manipulation.
For compliance officers, this development is relevant because the EU AI Act obliges providers of AI systems to label synthetically generated content. Anthropic explicitly positions the watermarking feature in Claude as an implementation of this regulatory requirement. Companies that use Claude in products or internal processes should check whether and how this labeling feeds into their own documentation and evidentiary obligations toward supervisory authorities.
From a compliance perspective, it remains to be clarified how robust a technical watermark is as sole proof of AI authorship within the meaning of the AI Act, especially when texts are further processed, translated, or converted into other formats. Companies should incorporate Anthropic’s documentation on the watermarking technique into their risk assessment for the use of generative AI systems and monitor whether comparable approaches are adopted by other providers or recognized by supervisory authorities as a reference standard.
Source: www.golem.de · Published August 17, 2026
Lumi AI News — AI-assisted curation in accordance with Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.