Updated
Updated · Forbes · Aug 12
Anthropic Watermarks Claude Text Globally as EU Rules Threaten €15 Million Fines
Updated
Updated · Forbes · Aug 12

Anthropic Watermarks Claude Text Globally as EU Rules Threaten €15 Million Fines

3 articles · Updated · Forbes · Aug 12

Summary

  • Claude models launched from Aug. 2 now embed invisible watermarks in all generated text, while supported files such as PNGs, JPGs and SVGs carry signed provenance metadata.
  • Article 50 of the EU AI Act took effect on Aug. 2, requiring machine-readable labels on generative AI outputs; Anthropic said it is applying the system worldwide across its API, Claude Code and cloud deployments.
  • Anthropic said the text mark survives copying and some editing because it biases word choice statistically, and it plans detection tools for users and third parties to check text and files.
  • The company also warned detection is imperfect: older models, short passages, heavy paraphrasing and stripped file metadata may evade checks, while Claude-polished human writing could still test positive.
  • User backlash has centered on false-accusation risks, missing accuracy thresholds and no published dispute process, adding pressure as rivals including OpenAI weigh whether to watermark text.

Insights

If invisible AI watermarks are easily erased by heavy editing, are they truly stopping malicious actors or just punishing honest users?
Where is the line drawn between human authorship and AI generation when a simple synonym swap triggers a hidden tracking code?