Home
Anthropic adds invisible watermarks to AI text; hackers and developers swiftly launch de-watermarking tool

Anthropic recently embedded an "imperceptible watermark" into select Claude models to facilitate the identification of AI-generated content. This technique relies on statistical patterns derived from Claude’s distinct word choices, persisting even when users copy and paste the output. The company asserts this initiative aligns with its obligations under the EU AI Act.
Nevertheless, the move has ignited significant debate within the tech community. Critics worry that basic tasks like proofreading, translation, or summarization will inadvertently propagate these hidden markers in final outputs. Some users have even canceled subscriptions in protest, prompting a surge in search interest and the rapid development of tools designed to strip AI markers from text.
In response to user backlash, developers swiftly introduced solutions. Guillaume Meyer, founder of the AI e-commerce tool Memo, released an open-source project named Watermarks Remover shortly after Anthropic’s announcement. This tool eliminates hidden characters and metadata while rewriting text to preserve meaning, thereby disrupting the statistical patterns associated with watermarks. Developed in roughly five hours, the project garnered over 14,000 GitHub stars, though Meyer acknowledged it cannot guarantee complete watermark removal.
Meyer publicly endorsed content attribution but criticized the watermark technology for oversimplifying authorship into a binary classification. Similarly, Sabrina Ramonov, an AI education provider, launched a free browser-based Watermark Remover capable of stripping hidden markers from text, PDFs, Word files, web pages, images, and data. Meanwhile, Tokyo-based developer Ansh Aneja created a dedicated Claude watermark removal tool on the same day as Anthropic’s announcement, later releasing an open-source local version called MarkScrub, which saw rapid user growth within 24 hours.
From technical and legal standpoints, these AI watermark removal tools occupy a precarious gray area. Thibaud Gloaguen, a researcher at ETH Zurich, noted that these tools underscore the inherent limitations of watermarking technology, as rephrasing and other methods will always offer ways to bypass them.
Legally, the EU AI Act mandates that AI content marking systems remain resilient against common modifications and adversarial attacks, requiring assessments of potential threats from watermark removal. However, the act does not explicitly ban third parties from developing or using such tools. Legal experts warn that maliciously disguising AI content as human-written may violate Anthropic’s terms of service, which prohibit impersonation. As Anthropic prepares to launch a text detection API with its next-generation model and expand watermarking to older versions, the conflict between AI watermarks and de-watermarking tools is likely to intensify.
Related article
Kuaishou Q2 Revenue Hits 35.5 Billion Yuan; Kuaichuan AI Sales Surge Over 200%
Kuaishou Technology’s Q2 2026 financial results reveal total revenue of 35.5 billion yuan, driven by a 7.4% year-over-year increase in core commercial revenue. The company reported an adjusted net profit of 3.9 billion yuan, reflecting an 11% net pro
AI Assistant Yuanbao Boosts Image Understanding to Meet All Daily Needs
Tencent’s AI assistant, Yuanbao, has unveiled a comprehensive upgrade to its image recognition capabilities, introducing a new feature titled “Stay Calm, Take a Photo First.” This enhancement prioritizes precision in text extraction, nuanced detail a
Jiangxi launches first token factory in Nanchang High-tech Zone, hosting 300+ large models
Jiangxi Province has officially launched its first Token factory in the Nanchang High-tech Industrial Development Zone, marking a significant step in regional digital transformation. Designed as a core AI empowerment platform, this initiative deliver
Related Special Topic Recommendations
Comments (0)
0/500

Anthropic recently embedded an "imperceptible watermark" into select Claude models to facilitate the identification of AI-generated content. This technique relies on statistical patterns derived from Claude’s distinct word choices, persisting even when users copy and paste the output. The company asserts this initiative aligns with its obligations under the EU AI Act.
Nevertheless, the move has ignited significant debate within the tech community. Critics worry that basic tasks like proofreading, translation, or summarization will inadvertently propagate these hidden markers in final outputs. Some users have even canceled subscriptions in protest, prompting a surge in search interest and the rapid development of tools designed to strip AI markers from text.
In response to user backlash, developers swiftly introduced solutions. Guillaume Meyer, founder of the AI e-commerce tool Memo, released an open-source project named Watermarks Remover shortly after Anthropic’s announcement. This tool eliminates hidden characters and metadata while rewriting text to preserve meaning, thereby disrupting the statistical patterns associated with watermarks. Developed in roughly five hours, the project garnered over 14,000 GitHub stars, though Meyer acknowledged it cannot guarantee complete watermark removal.
Meyer publicly endorsed content attribution but criticized the watermark technology for oversimplifying authorship into a binary classification. Similarly, Sabrina Ramonov, an AI education provider, launched a free browser-based Watermark Remover capable of stripping hidden markers from text, PDFs, Word files, web pages, images, and data. Meanwhile, Tokyo-based developer Ansh Aneja created a dedicated Claude watermark removal tool on the same day as Anthropic’s announcement, later releasing an open-source local version called MarkScrub, which saw rapid user growth within 24 hours.
From technical and legal standpoints, these AI watermark removal tools occupy a precarious gray area. Thibaud Gloaguen, a researcher at ETH Zurich, noted that these tools underscore the inherent limitations of watermarking technology, as rephrasing and other methods will always offer ways to bypass them.
Legally, the EU AI Act mandates that AI content marking systems remain resilient against common modifications and adversarial attacks, requiring assessments of potential threats from watermark removal. However, the act does not explicitly ban third parties from developing or using such tools. Legal experts warn that maliciously disguising AI content as human-written may violate Anthropic’s terms of service, which prohibit impersonation. As Anthropic prepares to launch a text detection API with its next-generation model and expand watermarking to older versions, the conflict between AI watermarks and de-watermarking tools is likely to intensify.
Kuaishou Q2 Revenue Hits 35.5 Billion Yuan; Kuaichuan AI Sales Surge Over 200%
Kuaishou Technology’s Q2 2026 financial results reveal total revenue of 35.5 billion yuan, driven by a 7.4% year-over-year increase in core commercial revenue. The company reported an adjusted net profit of 3.9 billion yuan, reflecting an 11% net pro
AI Assistant Yuanbao Boosts Image Understanding to Meet All Daily Needs
Tencent’s AI assistant, Yuanbao, has unveiled a comprehensive upgrade to its image recognition capabilities, introducing a new feature titled “Stay Calm, Take a Photo First.” This enhancement prioritizes precision in text extraction, nuanced detail a
Jiangxi launches first token factory in Nanchang High-tech Zone, hosting 300+ large models
Jiangxi Province has officially launched its first Token factory in the Nanchang High-tech Industrial Development Zone, marking a significant step in regional digital transformation. Designed as a core AI empowerment platform, this initiative deliver











