Developers Quickly Crack Claude Text Watermark, Fueling Industry Debate
Anthropic embedded a statistical watermark in Claude-generated text as part of its effort to meet transparency expectations under the European Union’s AI Act. The technique alters patterns in word selection so automated systems can identify likely machine-written material without adding a visible label. Such tools are intended to support content provenance, copyright enforcement and misinformation controls, but researchers have long questioned whether text watermarks can survive paraphrasing, editing or deliberate evasion.
Within hours of the watermark becoming public, several developers released open-source tools that they said could detect and strip Claude’s statistical signature. The rapid workaround intensified debate over whether text watermarking can serve as a durable regulatory safeguard when users can remove it at little cost. The episode also raises practical questions for Anthropic and policymakers about platform accountability, ownership of AI-assisted work and the evidentiary value of labels that may disappear after routine rewriting.
All Coverage
1 original reportsThe Backstory
The history behind this eventAnthropic Adds Invisible Watermarks and Digital Signatures to Claude
Anthropic is adding provenance markers to Claude output as AI companies face mounting pressure to make synthetic content identifiable. The system combines machine-readable, invisible patterns in generated text with C2PA digital signatures for supported files, creating a record of origin without visibly altering the material. The move comes as relevant transparency obligations under the European Union’s AI Act took effect on Aug. 2, 2026, though watermarking remains imperfect because substantial rewriting can erase the signal.
Anthropic recently explained that its text watermark uses statistical patterns designed to preserve Claude’s output quality while leaving a machine-detectable record. The rollout prompted complaints and reported cancellations from some paying users, while developers quickly began testing ways to defeat it. One removal tool attracted more than 9,000 GitHub stars within three days. Critics also note that no broadly available public verifier can independently assess the text watermark’s effectiveness, underscoring the gap between provenance safeguards and durable detection.
Subscribe to Mark Radar Weekly
Every Friday, the week's strongest signals in your inbox. Unsubscribe anytime.
If you search news on Google, you can set Mark Radar as a preferred source—our coverage will show up more often in your results. Set as preferred source on Google →