Anthropic's decision to embed invisible, machine-readable watermarks in Claude-generated text to comply with the European Union's AI Act has already drawn a developer response. According to WIRED, developer Guillaume Meyer published an override within four hours of Anthropic confirming the watermarking, and his code has since gone viral on GitHub.
Override code spreads within hours
According to WIRED, Meyer's code has been bookmarked more than 20,000 times on X and has drawn more than 100 contributors, with many more incorporating it into their own projects. "Anthropic is embedding watermarks in its Claude texts … the issue is practically history just one day later," wrote one AI specialist. Meyer and others started investigating watermarking after Anthropic announced last week that Claude would adopt it to comply with the EU AI Act. Meyer told WIRED that some developers try to evade watermarking because they disagree with labeling all AI-generated content, while others, including himself, simply relish the technical challenge. Freelance content writers and social media creators have also contacted Meyer asking for assistance using the code, he said.
The case against watermarking
"I'm not against transparency, and I'm all for content attribution. I just think watermarking in itself is a really bad solution, because it has major drawbacks and risks." — Guillaume Meyer
Meyer told WIRED he is concerned about false positives and the possibility that watermarking may not distinguish between light or heavy use of AI. As a native French speaker, he often uses Claude and tools like Grammarly to edit his writing. Because Anthropic admits it can only generate a probability that text has been touched by Claude, Meyer says using the watermark as evidence could lead employers to unfairly reject candidates or overblown accusations of researchers using artificial intelligence.
How Claude's watermark works
Anthropic watermarks text by leaving a pattern in Claude's choice of words and phrases that is indiscernible to a human reader but detectable by a machine that knows how to look for it, according to WIRED. Because this influences Claude's output, some users worry it will degrade response quality, though Anthropic insists it won't. The technique, called SynthID, was developed by Google, which has used it since 2023. Computer scientist Scott Aaronson proposed a similar method while at OpenAI, but WIRED reported the firm never deployed it because it was worried watermarks would put customers off its product.
Bypass method and compliance timeline
Meyer's removal method uses a non-watermarking large language model to generate multiple rewrites, swapping in synonyms and slightly reorganizing content, WIRED reported. This relies on other LLMs that do not insert watermarks — a bet that may not be safe, since 190 organizations, including OpenAI, Microsoft, and Meta, have signed the EU's transparency code of practice. It remains to be seen how many labs will implement watermarks. Under the EU AI Act, new models released from August must include watermarks, and existing models must have them integrated by December.
The EU rules, which came in earlier this month, require model providers like Anthropic and OpenAI to label synthetic audio, image, video, or text so a machine can detect it as AI-generated — or face fines of up to 3 percent of annual turnover. While providers cannot market circumvention tools, WIRED noted there is no legal restriction on independent tools.
| EU AI Act watermark requirements | Details |
|---|---|
| New models | Must include watermarks from August |
| Existing models | Must have watermarks integrated by December |
| Non-compliance penalties | Fines up to 3% of annual turnover |
| Transparency code signatories | 190 organizations, including OpenAI, Microsoft, Meta |
Wayne Pan, chief technology officer and cofounder at a Silicon Valley–based company, told WIRED that while there is no certainty Meyer's tool works until Anthropic releases the software it uses to detect a watermark, understanding the basic SynthID-text approach underpinning Claude's watermarking makes developers fairly sure the method works.
For enterprises deploying Claude, the episode underscores unresolved questions about detection reliability and the pace of circumvention. Until Anthropic ships its detector, organizations cannot validate whether text output carries a reliable watermark.
Sources:
