Claude now watermarks its output, and you cannot switch it off

Text produced by recent Claude models now carries a hidden, machine-readable mark, and there is no setting to turn it off. Anthropic applies the watermark at the model level, which means it is present regardless of which product the text came out of.

What is marked

Anthropic says Claude models launched on or after 2 August 2026 support marking at launch, and that work to bring older models into line is in progress. The surfaces it lists are the Claude API, the Claude apps, Claude Code, Claude Cowork and Claude Tag, plus Claude offered through AWS, Google Cloud and Microsoft Foundry. Generated text carries the watermark itself. Generated files in .svg, .png and .jpg formats instead get signed provenance metadata under the C2PA standard.

The trigger is the transparency requirement in the EU AI Act, which applied from 2 August 2026. Because the marking is done by the model rather than by a regional switch, it is not confined to European users.

How solid is the mark

Anthropic's own description is unusually measured. The watermark survives copying and pasting and may survive some editing, but the company states that heavily edited, paraphrased or translated text can lose detectability. It flags the converse too: no mark detected does not establish that a piece of text was not AI-generated.

Detection is the piece that has not landed. Anthropic says it will publish details of detection mechanisms in forthcoming technical documentation, so as things stand nobody outside the company can test a document. The marking is running now and the readers arrive later, which means work published this month becomes checkable retrospectively once they do.

What to do

• Read your client contracts. If you deliver copy under terms that restrict AI assistance or require it to be disclosed, the evidence question has stopped being hypothetical.
• Do not treat a light edit as removal. Anthropic only claims the mark degrades under heavy rewriting, and you have no way to check which side of that line your draft sits on.
• Change the disclosure, not the workflow. Saying up front that you draft with AI costs far less than being asked about it after the fact.
• Note this is not Anthropic-specific. The same EU rule applies to the other large model providers, so treat single-vendor switching as no answer at all.