In brief: Anthropic describes the watermarking method for Claude texts, which serves as a technical response to the labeling obligations of the EU AI Act, but according to the company cannot be fully protected against removal.
Anthropic has disclosed how the text watermark implemented in Claude works technically and how robust it is against circumvention attempts. The background is the labeling obligations for AI-generated content as provided for by the EU AI Act.
In a technical explanation, Anthropic set out the method by which texts generated by Claude are watermarked. The company describes both the origin of the method used and its practical limitations: according to Golem.de, the watermark can only be removed with considerable effort, but complete circumvention is not ruled out.
For compliance officers, the publication is relevant because the EU AI Act obliges providers of AI systems to label synthetically generated content. Watermarking methods such as the one described by Anthropic are considered one possible technical implementation of this labeling obligation. The transparency regarding the functioning and weaknesses of the method allows companies to better assess the AI technology they use with respect to regulatory requirements.
In practice, this means that organizations using Claude or further processing its outputs should examine to what extent the existing labeling is sufficient for their own documentation obligations under the AI Act. Since, according to Anthropic, the watermark is difficult but not impossible to remove, it should not serve as the sole basis for proof of origin of content but rather be regarded as a supplementary measure.
Source: www.golem.de · Published August 17, 2026
Lumi AI News — AI-assisted curation pursuant to Art. 50 EU AI Act. Paraphrasing and classification by Lumi News Pipeline v1.8.3.