Anthropic is now marking the content its Claude models produce. Since August 2, the company has been weaving an invisible watermark into generated text and attaching signed metadata labels to supported image files, in a push for more transparency around AI-made content.
How the watermark works
The text watermark is embedded at the model level, not as a hidden tag or piece of metadata. Anthropic says it travels with copied text and may survive some editing, which makes it harder to strip out by accident. The company has not released the full technical details of the mechanism, and it has not yet published a public detector tool, though it says one is coming.
"A watermark only helps test whether Claude might have produced or processed the content," Anthropic said. "It doesn't say anything about ownership or authorship, and doesn't change a user's rights."
C2PA labels for images and files
For files such as PNG, JPG and SVG images, Claude attaches a content credential: a small, cryptographically signed note in the file's metadata saying the file was made or processed with Claude. The label uses the C2PA open standard, the same system used by camera manufacturers and photo-editing software to record where an image came from. Any C2PA-aware tool can read it, and the signed record makes tampering detectable.
Reaction from writers and developers
The rollout has drawn a mixed response. Anthropic says the change follows its commitment to the European Commission's code of practice on transparency of AI-generated content under the EU AI Act. Some writers and developers welcomed the provenance labels but objected to the hidden text watermark, arguing that it changes written work without the user's clear consent and that only Anthropic, which holds the secret keys, can verify it for now. Regulators and fact-checkers, meanwhile, see the markings as a useful tool for tracing where content came from.