Developers race to break Anthropic's AI watermark hours after launch

Anthropic's plan to automatically embed invisible watermarks in Claude-generated content hit an immediate snag when independent developers published working workarounds just hours after the company announced the feature. One programmer released removal code that has since accumulated over 20,000 bookmarks on X and attracted contributions from more than 100 developers.

The rapid response#

Developer Guillaume Meyer shared a removal tool within four hours of Anthropic's watermark announcement, sparking widespread interest in the developer community. The technique uses non-watermarked language models to rewrite Claude's output by substituting synonyms and reorganizing sentences. Meyer's code has been integrated into numerous projects and shared across platforms, suggesting developers view watermark circumvention as an achievable technical challenge rather than a permanent barrier.

Other engineers have independently developed competing solutions. Software developer Erik Hughes created a tool that removes invisible characters, reorders sentences, and swaps vocabulary. Oxford researcher Leon Chlon noted that translation workflows—moving text into Arabic and back to English—can strip watermarks due to the language's distinct semantics.

Why the EU pushed watermarks#

Anthropic implemented the invisible marking system to comply with the European Union's AI Act, which took effect earlier this month. The regulation requires AI providers to label synthetic audio, images, videos, and text in machine-readable ways or face fines of up to 3 percent of annual turnover. Anthropic says this transparency helps people identify AI-generated content, though the company cannot market circumvention tools itself.

Concerns about the approach itself#

Developers challenging the watermark system often argue the solution creates problems it doesn't solve. Meyer raised the risk of false positives and the watermark's inability to distinguish between light editing (using Claude to polish grammar) and full content generation. Since Anthropic itself admits watermarks only yield probability estimates, using them as hard proof could unfairly penalize job candidates or trigger false accusations of AI misuse.

Meyer, a native French speaker who regularly uses Claude alongside tools like Grammarly, worries the invisible marking could degrade response quality—though Anthropic denies this will occur. The company says the watermark pattern emerges naturally in word and phrase selection without affecting meaning or readability.