Within hours of Anthropic confirming that Claude models would embed invisible, machine-readable watermarks into all AI-generated content, developer Guillaume Meyer published code to strip them. His tool went viral on GitHub, was bookmarked over 20,000 times on X, and attracted more than 100 contributors, with many others building on it.
The move follows Anthropic's announcement last week that Claude would adopt watermarking to comply with the European Union's AI Act. New rules, effective this month, require model providers like Anthropic and OpenAI to label synthetic audio, image, video, or text so machines can detect AI-generated content, or face fines up to 3 percent of annual turnover. While providers can't market circumvention tools, there's no legal ban on independent ones.
Meyer says some developers oppose labeling all AI-generated content, while others, including himself, enjoy the technical challenge. Freelance writers and social media creators have also reached out for help using the code. He argues watermarking has major drawbacks, including false positives and failure to distinguish between light and heavy AI use. As a native French speaker, he often uses Claude and Grammarly for editing, and worries employers might unfairly reject candidates based on watermark detection probability.
Anthropic embeds watermarks by leaving a pattern in Claude's word choices, invisible to humans but detectable by machines. Some users fear this degrades response quality, though Anthropic insists it won't. The technique, called SynthID, was developed by Google and used since 2023. Computer scientist Scott Aaronson proposed a similar method at OpenAI, but the firm never deployed it over fears it would deter customers.
Meyer's removal method uses a non-watermarking large language model to generate rewrites, swapping synonyms and reorganizing content. This relies on other models without watermarks, though 190 organizations—including OpenAI, Microsoft, and Meta—have signed the EU's transparency code. New models must include watermarks from August, and existing ones by December.
Other developers have built their own tools: Erik Hughes created one in 15 minutes that removes invisible characters and reorders sentences. Leon Chlon, a Visiting Fellow at Oxford, suggests condensing responses or translating into Arabic and back. Anthropic acknowledged that heavily edited, paraphrased, or translated content might not carry a watermark.
A spokesperson for Anthropic said: "We're adding marking to Claude's output to comply with the EU AI Act... It's hard to identify AI-generated text, and this gives people better tools for identification." The firm plans to release a text-detection API soon, which will let developers test if their removal methods work. Haimaker's Wayne Pan, who integrated Meyer's tool, doubts any watermark can withstand everything: "I don't think you can ever have a watermark that will withstand everything."
Source: Wired