Updated
Updated · Anthropic · Aug 13
Claude to Watermark AI Text Globally for EU AI Act, Covering Models Launched Before Aug. 2
Updated
Updated · Anthropic · Aug 13

Claude to Watermark AI Text Globally for EU AI Act, Covering Models Launched Before Aug. 2

3 articles · Updated · Anthropic · Aug 13

Summary

  • Anthropic said future Claude models will embed an undetectable statistical watermark in generated text, and older models launched before Aug. 2, 2026 will be updated over the coming months.
  • The watermark changes only the randomness behind low-stakes word choices, letting holders of a key estimate whether Claude likely helped write a passage without adding hidden characters, extra tokens, latency or cost.
  • Anthropic said the mark carries no user, organization or chat data, works poorly on short passages, and is weaker in factual text, proofreading and code where exact wording leaves little room for variation.
  • A detection API for text is coming soon, while Claude-generated images and files will use C2PA metadata credentials instead of text-style watermarking.
  • The rollout follows the EU AI Act and a July 2026 transparency code signed by Anthropic and about 190 others; Anthropic is applying the system globally because it cannot yet limit it by region.

Insights

Anthropic's invisible watermark catches AI text, but what happens when human writers subconsciously mimic these exact statistical patterns?
Can a simple paraphrasing tool completely erase Claude's new statistical watermark, rendering Anthropic’s major transparency update entirely useless?