Within four hours of Anthropic confirming that its Claude models would embed invisible watermarks into all AI-generated content globally, developer Guillaume Meyer published code to strip them out. His watermark removal tool has since been bookmarked more than 20,000 times on X and attracted over 100 contributors on GitHub, according to a report from WIRED. The swift workarounds raise questions about whether technical compliance with the European Union's new AI Act can actually achieve its transparency goals.
Anthropic announced the watermarking last week to comply with EU rules that took effect earlier this month, which require model providers to label synthetic text, audio, images, and video so machines can detect AI-generated material or face penalties up to 3 percent of yearly revenue. The watermarking technique, called SynthID and developed by Google, works by creating a pattern in Claude's word and phrase choices that humans can't perceive but machines trained to spot it can recognize. Google has used the same method to watermark its own AI-generated content since 2023. While the regulations prohibit providers from marketing circumvention tools, no legal restriction exists on independent developers building them. The rules mandate that new models released from August must include watermarks, and existing models must integrate them by December, though 190 organizations—including OpenAI, Microsoft, and Meta—have signed the EU's transparency code of practice, making it unclear how many will actually deploy watermarking.
Meyer told the outlet he's not opposed to transparency or attribution, but considers watermarking "a really bad solution" because of significant drawbacks and risks. He expressed concern about false positives and the technology's inability to distinguish between light and heavy AI use—particularly relevant since he uses Claude and tools like Grammarly to edit his writing as a native French speaker. Anthropic itself acknowledged that heavily edited, paraphrased, or translated content might not retain a watermark. The company told WIRED it's adding marking to comply with EU requirements and that "it's hard to identify AI-generated text, and this gives people better tools for identification," noting the watermark doesn't change the meaning, quality, or readability of responses.
Meyer's removal method relies on using a non-watermarking large language model to create multiple rewrites, substituting synonyms and slightly rearranging content. Other developers quickly built their own tools: software engineer Erik Hughes spent 15 minutes creating a tool with Claude that removes invisible and look-alike characters, reorders sentences within paragraphs, and swaps words for synonyms. Leon Chlon, a Visiting Fellow at the University of Oxford, says watermarks can be removed by condensing Claude's response, translating it into Arabic—which has very different semantics compared to English—and translating it back. Wayne Pan, chief technology officer and cofounder at Silicon Valley-based AI startup Haimaker, incorporated Meyer's open-source tool into his platform because he opposed Claude watermarking content that's only been lightly edited and disagreed with making the watermark invisible to users.
Anthropic says it's developing watermark detection for text and plans to release a tool soon, at which point developers will finally see whether their removal methods actually work. The company is also continuing to improve the watermarking system. Pan believes Anthropic wanted to demonstrate good faith compliance, but doubts any watermark can withstand all circumvention attempts. Meyer says freelance content writers and social media creators have contacted him for help using the code, while some are trying to evade watermarking because they disagree that all AI-generated content should be labeled, though he and others say they're motivated by the technical challenge itself. The fundamental tension between regulatory mandates and technical feasibility may force policymakers to reconsider whether machine-readable signatures can realistically enforce transparency when motivated developers can reverse-engineer and neutralize them faster than companies can deploy them.