Anthropic to Watermark Claude Text by December for EU Rules as Critics Question Quality Impact
Updated
Updated · The Guardian · Aug 17
Anthropic to Watermark Claude Text by December for EU Rules as Critics Question Quality Impact
3 articles · Updated · The Guardian · Aug 17
Summary
Anthropic said Claude will begin embedding undetectable text watermarks before December, altering low-stakes word choices so only Anthropic or key holders can later detect the pattern.
The change is meant to satisfy an EU rule requiring AI-generated text to carry watermarks, with Anthropic saying the system works at the model’s granular random generation layer.
John Gruber argued those constraints could degrade prose by steering Claude away from its best wording, but UCL professor Steven Murdoch said users would probably notice no difference because the model already relies on randomness.
That watermarking could also help identify chatbot-written work passed off by students or professionals and reduce the risk of models training on AI-generated text, a feedback loop linked to model collapse.