Anthropic extends textual watermarking to four more Claude models to meet the December deadline tied to the EU AI Act, The Decoder reports.
On the older Claude Opus 4.6 and Sonnet 4.6, developers can bypass the watermark by changing one sampling parameter.
For companies using these models, this means accounting for differences between versions when verifying the origin of generated text. The provided materials do not specify how often the bypass occurs, regulatory reactions, or Anthropic's plans to fix the issue.
editorial commentary
Why it matters
A likely consequence is that users and companies will have to separately verify the behavior of different Claude versions when complying with EU requirements. The next signal to watch is a clarification from Anthropic about a technical fix or regulators' position in Europe. Substantial uncertainty remains: the source does not specify the scale of the bypass and does not contain a primary statement from Anthropic.