Developers Create Tools to Bypass Claude's New AI Text Watermarks

Developers have already released tools designed to remove invisible watermarks from text generated by Anthropic's AI model, Claude. Shortly after Anthropic announced the implementation of these watermarks to comply with the European Union's AI Act, programmer Guillaume Meyer published code that has since gained significant traction on GitHub and X. The tool works by using a separate large language model to rewrite the text, swapping synonyms and reorganizing phrases to break the statistical patterns used for detection.

Anthropic's watermarking system is based on Google DeepMind's SynthID-Text technology. This method guides the AI's choice of words to create a machine-readable signature that remains indiscernible to human readers. While Anthropic claims that internal tests show no drop in output quality, some users express concern that the requirement to follow these patterns could degrade the quality of the content.

Critics of the technology, including Meyer, argue that the system is prone to false positives. They express concern that employers, universities, or recruiters might use these probability-based detectors as factual evidence to unfairly penalize individuals who use AI for minor editing or those who do not use AI at all.

The EU AI Act requires providers to label synthetic content or face fines of up to 3 percent of annual turnover. While approximately 190 organizations have signed a transparency code of practice, and watermarking is expected to become an industry standard by the end of the year, independent tools to circumvent these markers remain legal.