Anthropic Embeds Invisible Text Watermarks in Claude Outputs Globally

eyishazyer · x · 2026-08-11

Models launched after August 2 embed an invisible watermark directly into Claude's text globally, not just for EU users complying with regulations. Images, on the other hand, receive C2PA provenance metadata.

The author highlights that this watermark serves only as a probabilistic hint, not definitive proof. Heavy edits, paraphrasing, or translation can strip the text watermark completely, and short passages might lack the signal entirely. The real risk is that schools or employers might treat "watermark detected" as absolute proof of AI authorship, ignoring Anthropic's explicit documentation stating it shouldn't be used as such.

Related event: Anthropic embeds invisible watermarks in all Claude outputs, responding to EU AI Act(49 posts)→

Original post →

More from Models

Models channel →