Coders bypass Claude's AI watermarks in hours

Within four hours of Anthropic announcing global invisible watermarking for its Claude models to comply with the EU AI Act, developer Guillaume Meyer released an open-source script that removes these watermarks. The tool, now bookmarked over 20,000 times on X and drawing contributions from more than 100 developers, has gone viral as users seek to bypass labeling requirements without technically violating the EU’s rules—since the restrictions apply to providers, not independent tools.

The watermarking technique, called SynthID (developed by Google and adopted by Anthropic), embeds subtle patterns in word choices that are detectable by machines but invisible to humans. Critics like Meyer argue that such watermarks risk false positives, fail to distinguish between light and heavy AI use, and could unfairly penalize users—such as non-native speakers relying on AI for editing. Anthropic insists watermarking won’t degrade output quality, but users remain unconvinced. Meanwhile, Meyer’s approach leverages other non-watermarking LLMs to rewrite content, swapping synonyms and reorganizing text to strip detectable patterns—though this strategy may become riskier as major providers like OpenAI, Microsoft, and Meta prepare to roll out their own watermarks by August and December 2024.