AI Chat Daily
Important
AnthropicSafetyRegulationAnthropic Claude Watermarks Broken Within Hours of Launch
August 20, 20263 min read
A developer published a working override for Anthropic’s new text watermarking system within four hours of its wider rollout, drawing tens of thousands of bookmarks and exposing the practical limits of current AI watermarking approaches.
Why it matters
The rapid bypass raises questions about whether EU-mandated AI content labeling can be made robust enough to be useful in the real world.
Anthropic’s newly rolled-out text watermarking system for Claude was broken within roughly four hours of broader availability. Developer Guillaume Meyer published an override that quickly went viral on GitHub, collecting more than 20,000 bookmarks on X and attracting over 100 contributors.
The watermarking was introduced to meet EU transparency requirements by embedding machine-readable signals in generated text. The speed of the bypass has intensified debate about whether current statistical or cryptographic watermarking techniques can survive determined reverse-engineering once models are widely available.
The episode is likely to influence how regulators and labs think about the practical enforceability of AI content labeling rules.