Coders find workarounds to Anthropic’s invisible watermarks within hours of launch

Developers began releasing tools designed to remove Claude’s invisible AI watermarks within hours of Anthropic confirming their global rollout, WIRED reported.

Developer Guillaume Meyer published an open source workaround roughly four hours after Anthropic confirmed that supported Claude models would embed machine readable watermarks into generated text.

The project has since spread across GitHub and social media, attracting more than 100 contributors and over 20,000 bookmarks on X, according to WIRED.

Anthropic introduced the watermarking system as part of its compliance with transparency requirements under the European Union’s AI Act.

The rules require AI providers to make synthetic text, audio, images, and video machine detectable, with violations potentially resulting in fines of up to 3% of annual turnover.

Claude’s system uses Google’s SynthID technology, which creates an invisible statistical pattern through the model’s choice of words and phrases. The pattern is intended to be detectable by specialized software without changing the apparent meaning or readability of the text.

Meyer’s approach attempts to disrupt that pattern by using another large language model to rewrite passages, replace words with synonyms, and reorganize portions of the text.

Other developers have proposed similar approaches involving sentence rearrangement, paraphrasing, condensation, and translation.

Anthropic has acknowledged that heavily edited, paraphrased, or translated content may no longer retain a detectable watermark.

The company told WIRED that the technology is intended to improve identification of AI generated text and said it plans to release a text detection API.

Anthropic has not yet released the detection software required to independently determine how reliably the removal techniques defeat its watermarking system.