SciTech Pulse
AI Models

Anthropic to Watermark All Content Generated by Claude

Anthropic says all text, images and other output generated by its Claude AI models will now carry watermarks, a move it says complies with EU transparency rules while rolling out worldwide.

Step by step

  1. 1

    Model picks each next word

  2. 2

    Secret key steers the choice

  3. 3

    Small biases form a pattern

  4. 4

    Detector with key checks pattern

Roughly three-quarters of newly published web pages contain some machine-written text, according to a 2025 analysis by Ahrefs covering about 900,000 pages. Under Article 50(2) of the European Union's AI Act, providers of generative AI systems must make generated content detectable as artificially generated. Anthropic, the company behind the Claude AI models, has signed the EU's Code of Practice on Transparency of AI-Generated Content, one of roughly 190 signatories, and says it will watermark Claude's output across its ecosystem.

Text generated by Claude will carry an invisible statistical watermark, while supported image formats such as PNG, JPG and SVG receive signed metadata noting Claude's involvement, Anthropic said. Marking applies across the Claude Platform application programming interface (API), the Claude apps, Claude Code and third-party products built on Claude's API, since the watermark rides inside the output itself. Models launched on or after August 2, 2026 support marking at launch, with older models gaining it over coming months. Though the requirement is European, Anthropic is applying the watermark worldwide, since it has no durable way to limit it by region.

For text, Anthropic uses a version of Google DeepMind's SynthID-Text method. A language model normally picks each next word at random from several equally plausible candidates; a secret key and preceding text steer that choice instead, and across many words these small biases form a pattern a detector with the key can examine. Anthropic said the method adds no extra characters or tokens and does not change output quality or identify a user. For image files, Claude attaches a cryptographically signed credential to the file's metadata under the open C2PA standard, without altering the image itself.

Anthropic said a separate watermark-detection API is coming but not yet finalized. It called the marks signals, not proof, saying it "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this,'" and that a complete rewrite replacing every word would remove the text watermark entirely.

Terms explained

The story so far

  1. Anthropic details how it will hide watermarks in Claude's writing to meet EU AI Act
  2. Anthropic to Watermark All Content Generated by Claude
#anthropic#claude#watermarking#ai-content#eu-ai-act#c2pa
Rate this story

Related stories