Coders Say They Already Found Workarounds to Claude’s Invisible Watermarks
wired ·
Anthropic recently announced it would add invisible, machine-readable watermarks to Claude-generated content to comply with EU AI Act rules. Within hours, developer Guillaume Meyer published a code override to remove these watermarks, which went viral and attracted many contributors. Meyer and others object to watermarking because of concerns about false positives, quality degradation, and the inability to distinguish light versus heavy AI use. The EU rules require model providers to label synthetic content or face fines, but they do not prohibit independent circumvention tools. Meyer's method uses another language model to rewrite and reorganize text, while other coders proposed methods like translating text into Arabic and back. Anthropic acknowledged that heavily edited or translated content might escape watermarking. The company plans to release a text-detection API and said the watermark does not change the meaning, quality, or readability of responses. Experts remain uncertain whether any watermark can be fully resistant to removal.
wired