Particle.news

Anthropic Is Adding Invisible Watermarks to Claude Outputs Worldwide

Required by the EU AI Act, the change raises questions about how reliably companies can prove AI involvement when detection methods remain private.

Overview

  • Anthropic began embedding an invisible, machine‑readable watermark in text from Claude models launched on or after Aug. 2, 2026, and is attaching digitally signed C2PA provenance to generated files.
  • The watermark is text‑native: the model nudges word choices to create a faint statistical pattern that travels with copied text and can persist through light edits.
  • Anthropic says a detected mark signals Claude processed the text but is not proof of sole authorship and that a missing mark does not prove human origin.
  • Researchers and builders published proof‑of‑concept scrubbers within days and experts warn heavy editing, format changes, screenshots, or reprocessing can remove or break the signal.
  • Users and developers have pushed back with subscription cancellations and legal and privacy concerns, and Anthropic says it will publish detection tools and a free API but has not released detection details or error rates yet.