Coders broke Claude's invisible watermarks in under 4 hours
developers published a tool to bypass anthropic's synthid watermarks in four hours. this exposes the fragility of eu regulations relying on technical detection methods to label ai content.
anthropic globally activated invisible, machine-readable watermarks via google's synthid to comply with the eu ai act, but developer guillaume meier published a working override in four hours. the tool, which uses non-watermarked llms to rewrite text and strip the patterns, gained over 100 contributors and 20,000 bookmarks online. meier and critics argue watermarking harms output quality and risks false positives that could unfairly penalize users using ai for grammar or translation. with the eu threatening fines up to 3% of turnover, this rapid circumvention suggests policy enforcement may rely on easily broken technical safeguards.