Within four hours of Anthropic confirming in August 2026 that its Claude AI models would embed invisible watermarks into AI-generated content, a developer had already published working code to remove them.
Developer Guillaume Meyer released his watermark-removal tool shortly after Anthropic’s announcement, and it has since gone viral on GitHub, accumulating more than 20,000 bookmarks on X and attracting over 100 contributors. The tool works by using a non-watermarking large language model to generate multiple rewrites of text, swapping synonyms and slightly reorganizing content to strip the embedded pattern.
Anthropic adopted watermarking to comply with the European Union’s AI Act, which came into effect earlier in August 2026. The regulation requires AI model providers — including Anthropic and OpenAI — to label synthetic audio, image, video, and text so it can be machine-detected as AI-generated, or face fines of up to 3 percent of annual turnover. The rules prohibit providers from marketing circumvention tools, but place no legal restriction on independent tools.
Anthropic’s watermarking uses a technique called SynthID, developed by Google and in use since 2023. It works by leaving an imperceptible pattern in Claude’s word and phrase choices that a trained machine can detect. Because the technique influences Claude’s outputs, some users have raised concerns about response quality, though Anthropic says quality will not be degraded.
Meyer told WIRED his motivation is partly technical curiosity and partly concern about the watermarking approach itself. He worries about false positives — situations where light AI use, such as grammar editing, could trigger a watermark flag and lead to unfair outcomes for job applicants or researchers. Anthropic has acknowledged the watermark can only generate a probability that text was Claude-generated, not a certainty.
Wayne Pan, cofounder and CTO of AI startup Haimaker, incorporated Meyer’s tool into his platform, citing similar concerns about invisible watermarks being applied to lightly edited content. The EU’s transparency code of practice has been signed by 190 organizations, including OpenAI, Microsoft, and Meta. New AI models must include watermarks from August, with existing models required to comply by December 2026.
Source: WIRED