Developers Build Tools to Strip Anthropic's Claude AI Watermarks

Developers Build Tools to Strip Anthropic's Claude AI Watermarks

Within days of Anthropic announcing imperceptible watermarks for text generated by its Claude AI models, software developers released open-source tools designed to remove them. The rapid response highlights a fundamental challenge in AI content attribution: technical safeguards can be circumvented as quickly as they're deployed.

Anthropic introduced the watermarking feature last week to comply with the European Union's AI Act. The statistical pattern embedded through Claude's word choices persists when text is copied and pasted. But the announcement triggered immediate backlash from users concerned that watermarks would tag their work even when they only used Claude for proofreading or translation. Some users canceled their subscriptions, according to Business Insider, which first reported the story.

Google Trends data shows US searches for "AI watermark remover" jumped 60% week-over-week following the announcement.

Why it matters

The cat-and-mouse dynamic between watermarking and removal tools reveals a deeper problem for AI governance. Regulations like the EU's AI Act require providers to mark synthetic content, but enforcement mechanisms remain unclear when third parties can easily strip those markers. For enterprises using AI tools, this creates compliance uncertainty—especially when legitimate editing workflows might trigger watermarks that employees then feel compelled to remove.

First removal tools appear in hours

Paris-based entrepreneur Guillaume Meyer released "Watermarks Remover," an open-source project that strips hidden characters and metadata while rewriting text to disrupt the statistical patterns carrying watermarks. Meyer built the first version in approximately five hours. The project has accumulated more than 14,000 GitHub stars, a popularity metric similar to social media likes, though it does not guarantee watermark removal.

"I am all for content attribution," Meyer told Business Insider. "I am against the watermarking technique, and that's a very significant distinction." He argued that watermarking "treats authorship as a binary thing," applying the same marker whether Claude generated entire documents or merely corrected grammar.

AI educator Sabrina Ramonov created a browser-based "Watermark Remover" claiming to clean hidden AI markers from text, PDFs, Word documents, web pages, images, and data files. Tokyo-based developer Ansh Aneja built a Claude-specific remover on the day of Anthropic's announcement, later releasing an open-source version called MarkScrub. Aneja reported the tool gained 8,500 users in one day, though Business Insider could not independently verify those figures.

Legal gray zone

The EU's AI Act requires providers to implement durable watermarks resistant to common alterations and adversarial actions. However, the regulation does not explicitly prohibit third parties from creating or distributing removal tools, according to Dmitri Roussinov, a senior lecturer in Computer and Information Sciences at the University of Strathclyde.

Konrad Kollnig, an assistant professor at Maastricht University's Law and Tech Lab, noted that neither the AI Act nor Anthropic's terms appear to ban making or sharing removal tools. Users could face problems if they misrepresent AI-generated work as human-made—Anthropic's usage policy prohibits presenting model results as human-generated.

Thibaud Gloaguen, a researcher at ETH Zurich's Secure, Reliable, and Intelligent Systems Lab, pointed to an inherent limitation: "There will always be ways to remove the watermark," including simply rewording entire texts.

Anthropic plans to release a text-detection API alongside its next model, though no date has been specified. Watermarks apply to models released on or after August 2, with plans to extend the feature to older models. The company did not respond to questions about watermark removal tools.

These details were first reported by Business Insider.