Anthropic says it will start embedding machine-readable marks in the text and images Claude produces, so that platforms and people can establish where a piece of content came from. None of it will be visible to a reader. The commitment appears on a new Claude support page, and the driver is European transparency law.
It is a pledge rather than a shipped feature. The AI Act's labelling obligations took effect on 2 August, with a four-month grace period for products that already existed at that point. Anthropic's position is that new Claude models will mark their output from the day they launch, while retrofitting the models already out there remains in progress.
Two techniques are involved, and they are not equally well understood. For images the company will use C2PA, the provenance metadata standard Adobe, OpenAI and Google have already adopted. That part is conventional and the tooling around it exists.
The text side is considerably thinner on specifics. Anthropic describes an imperceptible watermark woven into the words themselves, leaving meaning, quality and readability untouched, and does not name the system behind it. Two properties are claimed. Because the mark lives inside the text, it travels with the text when it is copied and pasted, and may survive a degree of editing. And it is applied at the model level, so it should not matter which surface the text emerged from — Claude itself, the API, Claude Code, Cowork, Tag, or the model reached through AWS, Google Cloud or Microsoft Foundry.
Detection is the piece still missing. Anthropic says it is building a way for users and third parties to check content for these marks, with details to follow in technical documentation. C2PA readers already exist, Google's Gemini among them, but whether any of them will read Claude's files has not been confirmed.
What is notable is that Anthropic is not overselling the result. The company states plainly that these systems are far from infallible, and that content carrying no detectable mark may still have come from a generative model. That caveat is doing real work. C2PA metadata is famously easy to strip and often gets stripped by accident when a file is uploaded to a platform, and there is no public evidence yet about how much editing the text watermark actually withstands.
Demand for this exists in places that rarely make headlines. Readers on AO3 have been assembling their own crude tests for whether a fanwork was written with Claude's help. A watermark applied at the model level would do that job far better, and across the entire web rather than one community — assuming it survives contact with it.
