Particle.news

Anthropic Adds Invisible Watermarks to Claude Output

The company says the marks satisfy EU transparency rules but detection tools and full technical details have not yet been released.

Overview

  • Anthropic has begun embedding an invisible, machine‑readable signal in new Claude model outputs worldwide and plans to roll the feature across the app, API, Claude Code and Cowork; the company says the change has been in place since August 2 for supported models.
  • For text Anthropic uses a SynthID‑style method that steers token choices to create a hidden pattern and for supported files and images it will attach signed provenance metadata using the C2PA standard.
  • Anthropic warns the mark only suggests Claude may have processed content and cannot prove sole authorship while extensive edits, format changes or re‑saving can weaken or remove the signal and metadata.
  • The company has promised a public detection API and technical details but has not yet published them, leaving third‑party detection and verification tools unavailable and driving user concern.
  • Some users have publicly canceled subscriptions or switched providers over worries their work will be flagged, and experts warn the approach could be spoofed or stripped, raising practical, legal and contractual questions for professionals and educators.