Developers release tools to bypass Anthropic's Claude AI watermarks within hours of rollout

Within hours of Anthropic launching global invisible watermarks for Claude to comply with the EU AI Act, developers released open-source bypass tools. Developer Guillaume Meyer's workaround went viral on GitHub, drawing over 100 contributors. Critics and users express intense concern that the word-choice watermarking system, based on Google's SynthID, will degrade output quality and trigger false positives for people who only use Claude to edit their writing, leading some subscribers to cancel their accounts.

Watermark removal tools

• ▪Guillaume Meyer's watermark removal tool went viral on GitHub, drawing over 100 contributors, more than 14,000 GitHub stars, and over 20,000 bookmarks on X.
• ▪AI educator Sabrina Ramonov built a free browser-based tool designed to clean hidden watermarks from text, PDFs, Word documents, web pages, images, and data files.
• ▪Software engineer Erik Hughes built a tool using Claude in 15 minutes that removes watermarks by stripping invisible characters, reordering sentences, and swapping words for synonyms.
• ▪Tokyo-based developer Ansh Aneja released an open-source watermark removal tool called MarkScrub on the day of Anthropic's announcement.
• ▪Developer Guillaume Meyer published an open-source workaround tool to remove Claude's invisible watermarks approximately four hours after Anthropic confirmed their global rollout.

EU AI Act compliance

• ▪The European Union's AI Act requires AI providers to make synthetic text, audio, images, and video machine-detectable, with non-compliance fines of up to 3 percent of annual turnover.
• ▪While the European Union's AI Act restricts providers from marketing circumvention tools, it does not expressly prohibit independent third parties from developing or sharing watermark removal tools.
• ▪Anthropic introduced the global watermarking system for Claude to comply with transparency requirements under the European Union's AI Act.

Developer objections to watermarking

• ▪Tech commentator John Gruber argued on his blog that Anthropic's watermarking system would corrupt Claude's output, calling the factoring of external needs into user text generation offensive.
• ▪Guillaume Meyer objected to Anthropic's watermarking because it treats authorship as a binary thing, marking text regardless of whether Claude wrote it entirely or only assisted in editing.
• ▪Wayne Pan, cofounder of Haimaker, integrated Guillaume Meyer's removal tool because he disagreed with Claude watermarking lightly edited content and disliked that the watermark is invisible to users.

Claude watermark technical mechanism

• ▪Anthropic's watermarking system embeds an invisible statistical pattern in Claude's output by favoring specific word and phrase choices during text generation.
• ▪Anthropic acknowledged that heavily edited, paraphrased, or translated content might no longer carry a detectable watermark.
• ▪Anthropic's watermarking technique is based on Google's SynthID technology, which Google has used to watermark its own AI-generated content since 2023.
• ▪Anthropic plans to release a text-detection API to allow users to identify Claude-generated text, though independent verification of removal tools is limited until its release.

User subscription cancellations

• ▪Following Anthropic's watermark announcement, US Google Trends interest in the search term 'AI watermark remover' rose 60 percent week on week.
• ▪Some Claude subscribers cancelled their subscriptions in protest over Anthropic's implementation of the global watermarking feature.

False positive concerns

• ▪Anthropic stated that Claude's watermark only applies to words the model chooses, meaning a lightly edited document would contain too few changed words to trigger a detection mark.
• ▪Guillaume Meyer expressed concern that Claude's watermarks could cause false positives, leading to employers unfairly rejecting job candidates or researchers facing overblown accusations of AI use.

4 sources