Developers Create Bypass Tools Following Anthropic's AI Text Watermarking Release

Following Anthropic's implementation of statistical watermarking in its Claude models to comply with the EU AI Act, developers have quickly released open-source and web-based tools designed to remove these invisible markers. These tools, such as "Watermarks Remover" and "MarkScrub", aim to strip the statistical patterns embedded in the AI-generated text. This rapid response highlights the ongoing cat-and-mouse dynamic in AI governance, where technical safety measures are quickly met with bypass mechanisms. It also exposes a legal grey area under the EU AI Act, which mandates watermarking for AI providers but does not explicitly ban third parties from removing them. Anthropic's watermark is not based on hidden characters but rather on statistical patterns in word choice, which removal tools attempt to disrupt by rewriting or paraphrasing the text while preserving its meaning. While Anthropic plans to release a text detection API, experts note that public detectors make it easier for developers to test and refine their watermark removal tools.