Claude's Invisible Watermarks Bypassed Within Hours of Launch
Original: Coders Say They Already Found Workarounds to Claude’s Invisible Watermarks
Why This Matters
The rapid bypass highlights significant challenges in enforcing AI content labeling under the EU AI Act.
Within four hours of Anthropic announcing Claude's invisible watermark system to comply with the EU AI Act, developer Guillaume Meyer published a removal tool on GitHub that has since garnered 20,000+ bookmarks on X and over 100 contributors.
Anthropic last week announced that Claude models would globally embed invisible, machine-readable watermarks into AI-generated content to comply with the European Union's AI Act, which took effect earlier this month. The rules require model providers to label synthetic audio, image, video, or text so it can be machine-detected as AI-generated, with fines of up to 3% of annual turnover for non-compliance. Within four hours, French developer Guillaume Meyer published a watermark-removal tool on GitHub, which has since received over 20,000 bookmarks on X and attracted more than 100 contributors. Meyer stated his motivations were partly technical curiosity and partly concern over the reliability of watermarking — noting risks of false positives and that watermarks may not distinguish between light and heavy AI use. Freelance writers and social media creators have also sought his help. Anthropic uses Google's SynthID technique, which embeds watermarks by subtly influencing Claude's word and phrase choices in ways imperceptible to humans but detectable by machines. The company insists the watermarking does not degrade response quality. While providers cannot legally market circumvention tools, no legal restriction prevents independent developers from building them.