Anthropic says Claude will add invisible marks to AI text and images

In short: Anthropic says it will start adding hidden, computer-readable marks to text and images made with its Claude AI.

What happened

Anthropic published a new support page saying Claude-generated text will include embedded watermarks, and some generated files will include digitally signed “provenance metadata.” Provenance metadata is like a built-in label that records where a file came from, similar to a receipt attached to a photo.

The marks will be invisible to people reading the text or looking at the image. Anthropic says they are meant to help websites and other tools detect whether something was generated by Claude.

This is a commitment for the future, not something that switches on immediately for every Claude user. Anthropic says new Claude models will include these marks from day one, while support for existing models is still being worked on.

For images, Anthropic plans to use C2PA, a standard used by other companies to attach origin information to files. For text, Anthropic says it will weave an “imperceptible watermark” into the wording without changing the meaning, and it may still be detectable after copying, pasting, and some editing.

Anthropic says the marks will be applied globally across supported Claude products, including its API (a way for other apps to connect to Claude), and also when Claude is used through AWS, Google Cloud, or Microsoft Foundry.

Why it matters

As AI content spreads online, it is getting harder to know what was made by a person and what was made by a tool. Hidden labels could make it easier to filter, flag, or verify AI-made content, but they are not perfect. Anthropic notes that metadata like C2PA can sometimes be removed, and content without a detectable mark could still be AI-generated.

Source: The Verge AI