Coders Say They Already Found Workarounds to Claude’s Invisible Watermarks

Recent developments have seen coders unveil methods to bypass Claude's concealed watermarks, according to a recent report. GitHub saw their code circulate widely, garnering over 20,000 bookmarks and over 100 contributors. The controversy arose after Anthropic's announcement of embedding watermarks in Claude texts due to compliance with the European Union's AI Act.

While some developers oppose the watermarking, arguing it could degrade content quality or lead to unfair accusations, others enjoy the challenge of breaking the system. The EU's new rules mandate model providers to label synthetic content, unless they face fines up to 3 percent of their annual revenue. Despite Anthropic insisting the watermark will not affect Claude's response quality, concerns persist about false positives and the watermark's potential to flag legitimate AI use.

The watermark, known as SynthID, leaves a subtle pattern in the choice of words and phrases that machines can detect. To counter this, some developers have created tools that use different large language models to generate content without watermarks, though this approach relies on the availability and trustworthiness of these models.

Other methods include manipulating the text by reordering sentences, swapping synonyms, or translating it into a different language before translating it back. While Anthropic acknowledges that heavy editing might not leave a watermark, they remain committed to improving the watermarking system and expect to release a detection API soon.