Within four hours of Anthropic confirming that its Claude models would globally embed invisible, machine-readable watermarks, developer Guillaume Meyer had published an override. His code to remove these watermarks quickly went viral on GitHub and X, drawing over 100 contributors and highlighting the immediate friction between regulatory compliance and developer autonomy.

The EU AI Act Mandate

Anthropic’s decision to implement watermarking follows the European Union’s AI Act, which requires providers like Anthropic and OpenAI to label synthetic audio, image, video, or text. Failure to comply could result in fines of up to 3 percent of annual turnover. While the law prohibits providers from marketing circumvention tools, there is currently no legal restriction on independent, open-source projects that bypass these labels.

Methods of Circumvention

The watermarking technique, known as SynthID and originally developed by Google, leaves a specific pattern in the choice of words and phrases. It is indiscernible to human readers but detectable by specialized software. However, coders have already identified several ways to strip these markers:

  • Using non-watermarking LLMs to rewrite and reorganize content.
  • Swapping words for synonyms and reordering sentences within paragraphs.
  • Translating text into dialects with different semantics, such as Arabic, and then back into the original language.

Concerns Over Accuracy

Critics like Meyer argue that watermarking is a flawed solution. He expressed concerns regarding false positives and the inability of the system to distinguish between light editing and full AI generation. This poses a risk for professionals, such as researchers or job applicants, who might be unfairly accused of using AI when their work is only partially assisted by tools like Claude or Grammarly.

Anthropic maintains that the watermarks do not degrade the quality of responses and plans to release a detection API soon. This will allow users to verify for themselves whether the current circumvention methods are truly effective.