Claude will now mark AI-generated content - here's how it actually works

Anthropic has published the details of how Claude flags content it generates, tying the move to a legal commitment: the company signed the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content, and the marking system is how it’s meeting that obligation.

The short version: if you’re reading text or looking at an image that came out of Claude, there’s now a decent chance it carries an invisible signal saying so, though that signal is far from universal yet, and it’s not something you’ll ever consciously notice.

How the Marking Works

Claude uses two separate methods, according to the support documentation. Generated text gets an imperceptible watermark woven directly into it, it doesn’t change how the text reads, and because it’s embedded in the text itself, it travels along when the text is copied and pasted elsewhere, and can survive some editing. Files that Claude generates in supported formats, such as .svg, .png, or .jpg, get signed provenance metadata instead, following the C2PA standard already used across the industry to track where content came from.

The coverage is broad by design: marking applies across Claude Platform (the API), the Claude apps, Claude Code, Claude Cowork, and Claude Tag, and also carries over when Claude is accessed through AWS, Google Cloud, or Microsoft Foundry. Anthropic says it applies wherever Claude is offered worldwide, not just in the EU.

There’s an important catch, though. Marking only applies to Claude models launched on or after August 2, 2026, a date that’s already passed as of this writing. Anthropic says it’s working to extend marking to older models, but hasn’t said when that will land.

The Limits Worth Knowing

Anthropic is upfront that a detected mark isn’t proof of anything definitive. Claude is often used to proofread, translate, or summarize someone else’s work, so a mark can show up on content Claude didn’t originally write. And the reverse is just as true: heavily edited text, very short passages, or files that had their metadata stripped by a screenshot or format conversion may show no mark at all, even if Claude produced them.

Anthropic hasn’t yet released the tools for actually checking whether content carries one of these marks, that detection capability is still described as forthcoming, with technical documentation to follow.

Source