Anthropic's EU AI Act watermark stripped by open-source tool days after launch

Anthropic's watermarking initiative fulfills Article 50 of the EU AI Act, which from August 2026 requires AI-generated content to be machine-readable. The scheme embeds invisible watermarks directly into Claude-generated text without affecting meaning, and uses the open C2PA standard for digitally signed provenance metadata on images and other files to detect tampering.

The compliance win was short-lived. Paris developer Guillaume Meyer released a free open-source tool that strips the text watermark within days of rollout, per StartupFortune, immediately undercutting the provenance guarantee for text. The episode crystallizes a long-standing critique: text watermarks are inherently fragile because text can be paraphrased, re-tokenized, or run through a stripper without visible degradation, unlike cryptographically signed image metadata.

The community reaction is pointed. Developers frame the stripper's rapid appearance as evidence that watermarking is theater — satisfying a regulatory checkbox while providing little real-world provenance assurance. The contrast with the C2PA image path is instructive: signed metadata with tamper detection is far more robust than statistical text watermarking, suggesting Anthropic's text compliance is the weak link by design of the medium, not of the implementation.

What to watch is the regulatory response. Article 50 requires machine-readability, but if the mark is trivially removable, does compliance hinge on the provider's good-faith embedding or on the mark surviving in the wild? The stripper tool turns an abstract policy question into a live enforcement problem, and other labs shipping EU-AI-Act watermarks will be watching whether Brussels treats removability as Anthropic's problem or the remover's.