Developers publish tools to remove Claude’s EU-compliance watermarks within hours of launch

Anthropic confirmed that Claude models would globally embed invisible, machine-readable watermarks into AI-generated content to comply with the European Union’s AI Act, which took effect earlier in August. The watermarks are designed to be indiscernible to human readers but detectable by machines, using a technique called SynthID, originally developed by Google and deployed in its AI-generated content since 2023.

Within four hours of Anthropic’s announcement, developer Guillaume Meyer published open-source code to remove these watermarks from Claude-generated text. The repository on GitHub quickly gained traction, with more than 20,000 bookmarks on X and contributions from over 100 developers. Meyer stated that some users seek to evade watermarking due to philosophical objections to mandatory labeling, while others are motivated by technical challenges.

The EU AI Act requires model providers like Anthropic and OpenAI to label synthetic audio, image, video, or text as AI-generated so that detection tools can identify it, or face fines of up to 3% of annual turnover. However, the regulation does not prohibit the creation or use of independent circumvention tools, creating a potential enforcement gap.

Critics, including Meyer, argue that watermarking carries significant risks, such as false positives and an inability to distinguish between light and heavy AI use. Meyer, who frequently uses AI tools like Claude and Grammarly for editing, expressed concern that watermarks could lead to unfair rejections of job candidates or baseless accusations against researchers based on probabilistic detection.

Anthropic’s watermarking method influences Claude’s choice of words and phrases, though the company asserts it does not degrade the quality, meaning, or readability of responses. The technique works by embedding a pattern detectable only by machines that know how to look for it. Anthropic also plans to release a text-detection API to enable users to verify watermark presence themselves.

Other developers have created alternative removal tools. Software engineer Erik Hughes built a tool in 15 minutes using Claude to remove invisible characters, reorder sentences, and swap words for synonyms. Leon Chlon, a Visiting Fellow at the University of Oxford, described a method involving translation to a semantically distant language like Arabic and back-translation to strip watermarks. Anthropic acknowledged that heavily edited, paraphrased, or translated content may not retain the watermark.

Anthropic spokesperson stated that the company is adding marking to comply with the EU AI Act and that other labs are taking similar steps. They emphasized that the watermark does not change the meaning, quality, or readability of responses and that a detection tool is planned for release. Despite these assurances, experts like Wayne Pan, chief technology officer at Haimaker, suggest that no watermarking system can withstand all circumvention attempts, noting that 190 organizations, including OpenAI, Microsoft, and Meta, have signed the EU’s transparency code of practice.