Recently, the AI company Anthropic introduced an "imperceptible watermark" in some of its Claude models to make AI-generated text easier to identify. This watermark is created through a statistical pattern formed by Claude's unique word selection. Even if users directly copy and paste the content, the watermark will remain with it. Anthropic stated that this move is to fulfill its commitments under the EU AI Act.
However, this measure has sparked considerable controversy in the tech industry. Many users and professionals are concerned that even if users simply use Claude for basic text proofreading, translation, or summarization, this hidden mark will also be spread along with the final output. Previously, some users even reported to the media that they canceled their Claude subscription services because of this. In response, search热度 surged, and various tools designed to remove AI markers from text began to emerge.
Faced with user backlash and demand, several developers quickly took action to provide solutions. Guillaume Meyer, a tech entrepreneur from Paris and founder of the AI e-commerce tool Memo, launched an open-source project called Watermarks Remover a few days after Anthropic announced its plan. The tool removes hidden characters and metadata while rewriting the text to maintain the original meaning, thereby disrupting the statistical word selection patterns that may carry watermarks. The project was developed in about 5 hours and quickly gained over 14,000 stars on GitHub, although Meyer admitted that the tool cannot guarantee 100% complete removal of the watermark.
Meyer publicly stated that while he supports content attribution labeling, he opposes this irrational watermark technology that simplifies authorship into a black-or-white classification. Meanwhile, Sabrina Ramonov, who provides AI education services, said on social media that she developed a free browser-based Watermark Remover that claims to remove hidden AI markers from text, PDFs, Word documents, web pages, images, and data files. Ansh Aneja, a software developer based in Tokyo, also developed a dedicated watermark removal tool for Claude on the same day Anthropic announced the feature, and later released a local open-source version called MarkScrub, which reportedly saw a significant increase in users within a day.
From a technical and legal perspective, these AI watermark removal tools currently exist in a delicate gray area. Thibaud Gloaguen, a researcher at ETH Zurich, pointed out that these tools once again highlight the long-standing limitations of watermarking technology, as there will always be ways to remove watermarks through rephrasing and other methods.
In terms of legislation, the EU AI Act requires AI service providers' content marking systems to be resilient against common modifications and adversarial operations, and to assess the potential threats of removing or modifying watermarks. However, the act itself does not explicitly prohibit third parties from developing or attempting to remove AI watermarks. Nevertheless, legal experts point out that if users use such tools to maliciously disguise AI-generated content as human-created, they may violate Anthropic's terms of service, as its policies prohibit users from impersonating humans. As Anthropic plans to launch a text detection API alongside its next-generation model and gradually expand the watermark feature to older models, the conflict between AI watermarks and de-watermarking tools is expected to continue escalating in the future.
