Coders Say They Already Found Workarounds to Claude’s Invisible Watermarks (www.wired.com)

Developer Guillaume Meyer has swiftly created a workaround to Claude's newly implemented invisible watermarks, designed by Anthropic to ensure AI-generated content is machine-detectable. Within hours of the announcement, Meyer's code gained massive attention on GitHub and social media, reflecting a widespread resistance to the watermarking method. While some developers challenge the idea that all AI-generated content should be labeled, others are motivated by the technical puzzle it presents. This rapid response raises questions about the effectiveness of watermarking as a compliance measure for the EU AI Act, which mandates labeling of synthetic media under penalty of fines. The watermarking mechanism, known as SynthID, embeds a hidden pattern in the output that influences Claude's word choices, but critics fear it could lead to false positives and mischaracterization of lightly edited work. Meyer’s removal method uses a non-watermarking language model to rephrase text, while other engineers have developed similar tools through various methods, including translation and synonym replacement. As Anthropic works on a detection tool to evaluate these circumvention strategies, the AI/ML community is left contemplating the future of content attribution and the potential for continuous battles between watermarking technologies and creative evasion techniques.