Anthropic wants to make AI-generated text easier to identify. Some software developers are already building tools designed to make those labels disappear.
The AI company said last week the “imperceptible watermark” on text produced by certain Claude models would travel with it when copied and pasted because it’s a statistical pattern created through Claude’s word choices.
While some have welcomed the improved transparency that AI watermarks could offer, the move has also prompted concerns among techies that a label could follow work even if they only used Claude to proofread, translate, or summarize. Some users previously told Business Insider they have canceled their Claude subscriptions because of it.
The backlash has led others to seek out ways to get around it: Google Trends interest for “AI watermark remover” in the US is up 60% week-on-week. Some developers and entrepreneurs have taken matters into their own hands and created tools to remove AI watermarks from text and files.
Anthropic has said it’s introducing the watermark to meet its commitments to the European Union’s AI Act. The company did not respond to Business Insider’s questions about AI watermark removal tools.
‘The wrong answer to a real problem’
Guillaume Meyer, a Paris-based tech entrepreneur and the founder of Memo, an AI tool for e-commerce brands, released an open-source project called “Watermarks Remover” days after Anthropic announced its plans.
The tool strips hidden characters and metadata, then rewrites text to preserve its meaning, thereby disrupting the statistical word-choice patterns that can carry a watermark, Meyer said.
The first version took about five hours to build, he added. It has more than 14,000 GitHub stars, a measure of its popularity among developers that’s similar to “likes” on social media sites. 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.”
Meyer said his objection is that watermarking “treats authorship as a binary thing.” AI could leave a watermark when used to generate the text entirely, but also for lighter editing.
Anthropic said in a blog post that its watermark is not meant to establish authorship.
“I think it’s the wrong answer to a real problem,” Meyer said of efforts to watermark AI-generated text.
A new breed of AI watermark removers
Meyer is not alone in trying to take on AI watermarks.
Sabrina Ramonov, an AI educator for entrepreneurs, said in an X post this week that she created a free, browser-based “Watermark Remover” that says it can “clean hidden AI marks” from text, PDFs, Word documents, web pages, images, and data files.
Meanwhile, Ansh Aneja, a Tokyo-based software developer, said he built a Claude-focused watermark remover on the day Anthropic announced the feature. He later released a local, open-source version called MarkScrub.
In an X post last week, Aneja said the tool went from zero to 8,500 users in a day. Business Insider could not independently verify those figures.
These projects are still nascent, and Business Insider has not been able to independently assess their effectiveness. Other companies, including Google and OpenAI, also use watermarks for images, and there are scores of tools that promise to remove them.
Still, the tools illustrate a long-standing limit of watermarking, said Thibaud Gloaguen, a researcher at ETH Zurich’s Secure, Reliable, and Intelligent Systems Lab.
“There will always be ways to remove the watermark,” he told Business Insider, giving the example of rewording the entire text.
A gray area
Watermark removers sit in a legal gray area: the EU tells AI companies to add durable labels, but does not clearly set out rules for third parties seeking to remove them.
A European Commission spokesperson told Business Insider the bloc’s guidance requires providers’ marking systems to be resilient to common alterations and adversarial actions against them.
The spokesperson added the EU’s transparency code specifically identifies removal, regeneration, copying, and modification as threats that providers should assess.
Those are requirements for AI providers, however. The AI Act does not expressly ban third parties from trying to remove a watermark, said Dmitri Roussinov, a senior lecturer in Computer and Information Sciences at the University of Strathclyde.
If that same removal tool uses AI to regenerate text, though, its provider may have its own obligation under EU law to mark that newly AI-generated output, he told Business Insider.
Konrad Kollnig, an assistant professor at Maastricht University’s Law and Tech Lab, said neither the EU’s AI Act nor Anthropic’s terms appear to bar people from making or sharing removal tools.
Users could, however, face problems if they use one to misrepresent AI-generated work as human-made. Anthropic’s usage policy prohibits impersonating a human by presenting model results as human-generated.
That question could become more urgent when Anthropic releases its planned text-detection API alongside its next model. It has not specified a date yet. Watermarks will apply to new models released on or after August 2, and Anthropic is also planning to roll it out on older models.
“All these watermarking techniques have one big issue,” Kollnig said. “Whenever the watermarking detection tool is made public, anyone can check AI-generated contents — including text and videos — against that detector and can build tools to remove watermarks.”
Are you using an AI watermark remover at work? Contact this reporter via email at tspirlet@businessinsider.com or Signal at thibaultspirlet.40. Use a personal email address, a nonwork WiFi network, and a nonwork device; here’s our guide to sharing information securely.
Read the full article here
