Anthropic's Claude EU AI Act Watermark Stripped Days After Rollout

Anthropic began invisibly watermarking all Claude-generated text on August 2, 2026, a move made to comply with the EU AI Act's provenance requirements for AI-generated content. Within days, Paris-based developer Guillaume Meyer released a free open-source tool that reliably strips the watermark, reigniting a long-running debate over whether text watermarking can ever be enforceable.

Unlike image or audio watermarks, which can survive many transformations, statistical text watermarks are notoriously brittle: paraphrasing, re-tokenizing, or targeted removal tools can erase the signal while preserving meaning. Meyer's tool demonstrates that a determined user can defeat the mechanism with minimal effort, undercutting the compliance value regulators hoped to extract.

The episode highlights a structural tension in the EU AI Act's provenance provisions: mandating watermarks does not guarantee they persist through real-world use. It also echoes broader community unease—HN users flagged that Microsoft's Paint and Photos now invisibly watermark locally generated output with a GUID, raising privacy concerns alongside the enforceability questions. Reporting also noted quality complaints that watermarking may subtly degrade text output.

For Anthropic, the fast defeat is an awkward compliance optics problem rather than a technical failure of its core models, but it invites scrutiny over whether the industry's watermarking commitments are substantive or performative. Readers should watch how EU regulators respond to trivially strippable watermarks, whether Anthropic hardens the scheme, and whether the industry pivots toward cryptographic provenance or content-credential standards instead.