AI's Invisible Ink: Claude Now Watermarks Content

Anthropic has begun implementing machine-readable marks on content produced by its Claude AI models. This initiative aligns with the E.U. AI Act's transparency requirements, which mandate the labeling of AI-generated content and took effect on August 2. Supported Claude models launched on or after this date will include these markings from their release.

Text outputs from Claude will carry an embedded watermark designed to persist through copying and some editing. For image files like .png and .jpg, Anthropic is applying signed metadata using the C2PA open standard, a method also employed by Adobe and OpenAI. The company is working on enabling users and third parties to detect these marks, with technical documentation forthcoming.

However, Anthropic cautions that the system is not foolproof. A detected mark suggests potential processing by Claude but does not offer absolute certainty, as content might have undergone post-processing like proofreading or translation. Conversely, the absence of a mark does not confirm human origin, as factors like heavy editing or format conversions can obscure it.

Several other major tech companies, including Google, Meta, and Microsoft, have also pledged to adhere to the E.U.'s transparency code. The E.U. AI Act provides a four-month grace period for existing AI products to achieve compliance.