Coders Say They've Already Found Ways Around Claude's Invisible Watermarks
Within hours of Anthropic announcing invisible watermarks in Claude-generated text, developers had published tools to strip them out. The code has quickly gone viral, raising doubts about how durable the EU-driven labeling requirement will be.
Anthropic announced last week that its Claude models would begin embedding invisible, machine-readable watermarks in AI-generated text worldwide, in order to comply with the European Union's AI Act. Within about four hours, developer Guillaume Meyer had published code that removes these watermarks. His tool has since been bookmarked more than 20,000 times on X and attracted over 100 contributors on GitHub.
The EU rules, which took effect earlier this month, require providers such as Anthropic and OpenAI to label synthetic audio, image, video, or text so it can be machine-detected, with fines of up to 3 percent of annual turnover for noncompliance. While providers themselves are barred from marketing circumvention tools, there is no legal restriction on independently developed ones.
Meyer says some people are trying to evade the watermarking because they object to the principle that all AI-generated content should be labeled, while others, including himself, are simply drawn to the technical challenge. He says he supports content attribution but considers watermarking a flawed solution, citing risks of false positives and its inability to distinguish light editing from heavy AI use — which he worries could lead to unfair job rejections or unfounded accusations against researchers.
Anthropic's watermarking uses Google's SynthID technique, subtly influencing Claude's word and phrase choices in a way invisible to readers but detectable by machines, without affecting meaning or quality, the company says. Meyer's removal method relies on other large language models that don't insert watermarks to rewrite text, though this approach may become harder as 190 organizations, including OpenAI, Microsoft, and Meta, have signed the EU's transparency code of practice.
Other developers, including engineer Erik Hughes and Oxford visiting fellow Leon Chlon, built their own workarounds involving character removal, sentence reordering, synonym swaps, or translating text into another language and back. Anthropic has acknowledged that heavily edited, paraphrased, or translated content may not retain a detectable watermark. The company plans to release a detection tool that will let developers finally test whether their removal methods actually work.
