Anthropic wants AI-generated text to be easier to identify, but some software developers have already begun creating tools specifically designed to remove these markers.
The AI company said last week that text generated by some Claude models would contain an “imperceptible watermark.” Because this watermark is formed through a statistical pattern in Claude’s word choices, it remains even when users copy and paste the text.
Some people believe AI watermarking can improve content transparency and is a welcome initiative. However, the approach has also raised concerns among technology professionals: even when users merely use Claude to proofread, translate, or summarize text, the marker could be carried into the final content. Some users previously told Business Insider that they canceled their Claude subscriptions for this reason.
The backlash has also prompted some people to look for ways around the watermark. In the United States, Google Trends searches for “AI watermark remover” rose 60% month over month. Some developers and entrepreneurs have even started building their own tools to remove AI watermarks from text and files.
Anthropic said the watermark was introduced to fulfill its commitments under the European Union’s Artificial Intelligence Act (AI Act).
Guillaume Meyer, a Paris-based technology entrepreneur and founder of the AI e-commerce tool Memo, released an open-source project called “Watermarks Remover” just days after Anthropic announced the plan.
Meyer said the tool deletes hidden characters and metadata, then rewrites the text while preserving its original meaning, thereby disrupting the statistical word-choice patterns that may carry the watermark.
He said the first version of the tool took about 5 hours to develop. The project has now received more than 14,000 stars on GitHub, a measure of popularity for developer-community projects similar to “likes” on social media. However, the tool cannot guarantee that a watermark will be completely removed.
Meyer told Business Insider: “I fully support content attribution, but I oppose this watermarking technology. There is a very important distinction between the two.”
Meyer believes the problem is that watermarking technology “treats authorship as a black-and-white matter.” If AI is used to generate an entire passage of text, the system can leave a watermark, but a watermark may also remain even when users use AI only for light editing.
Anthropic, meanwhile, said in a blog post that its watermark was not intended to determine who authored the content.
Regarding the practice of adding watermarks to AI-generated text, Meyer said: “I think this is the wrong solution to a real problem.”
Meyer is not the only person trying to address the AI watermarking problem.
This week, Sabrina Ramonov, who provides AI education services for entrepreneurs, said on X that she had developed a free browser-based “Watermark Remover.” The tool claims to “clear hidden AI markers” from text, PDFs, Word documents, webpages, images, and data files.
Meanwhile, Tokyo-based software developer Ansh Aneja said he developed a watermark removal tool specifically for Claude on the day Anthropic announced the feature. He later released a local open-source version called MarkScrub.
Aneja said on X last week that the tool grew from 0 users to 8,500 users in a single day. However, Business Insider could not independently verify the figure.
These projects are still in their early stages, and Business Insider was also unable to independently assess their actual effectiveness. Other companies, including Google and OpenAI, likewise use watermarks in content such as images, and numerous tools already claim to be able to remove them.
However, Thibaud Gloaguen, a researcher at the Secure, Reliable and Intelligent Systems Laboratory at ETH Zurich, said these tools once again demonstrate the long-standing limitations of watermarking technology.
He told Business Insider: “There will always be ways to remove a watermark.” He explained that users could rephrase an entire piece of text, thereby disrupting the original watermark pattern.
AI watermark removal tools currently occupy a legal gray area. The EU requires AI companies to add persistent markings to generated content, but it does not clearly state whether third parties may actively remove those markings.
A European Commission spokesperson told Business Insider that the EU’s relevant guidelines require AI service providers’ content-marking systems to withstand common modification methods and adversarial operations targeting the marking systems.
The spokesperson also said that the EU’s transparency rules explicitly list removal, regeneration, copying, and modification as potential threats that AI service providers should assess.
However, these requirements apply to AI service providers. Dmitri Roussinov, a senior lecturer in the Department of Computer and Information Sciences at the University of Strathclyde, said the AI Act does not explicitly prohibit third parties from attempting to remove AI watermarks.
But he told Business Insider that if a watermark removal tool itself uses AI to regenerate text, its provider may also need to add the appropriate markings to the regenerated AI content under EU law.
Konrad Kollnig, an assistant professor at Maastricht University’s Law and Technology Lab, said that neither the EU’s AI Act nor Anthropic’s terms of service currently appears to prohibit people from developing or sharing AI watermark removal tools.
However, users could run into problems if they use such tools to disguise AI-generated content as human-created work. Anthropic’s usage policy prohibits users from impersonating humans by claiming that model-generated output was generated by humans.
The issue could become more prominent as Anthropic plans to launch a text-detection API alongside the release of its next-generation models. Anthropic has not yet announced a specific launch date. The watermark will apply to new models released on August 2 and afterward, while Anthropic also plans to gradually apply the feature to older models.
Kollnig said: “All of these watermarking technologies share one major problem: once watermark-detection tools are made public, anyone can use a detector to examine AI-generated content, including text and video, and then further develop tools to remove the watermarks.”