Claude now watermarks the text it writes for you
On 11 August 2026 Anthropic said that text produced by its newer Claude models carries a hidden watermark. A human reader cannot see it, a machine can read it, and it stays with the text when you copy it and paste it somewhere else. The marking covers every Claude model released on or after 2 August 2026. Older models are not marked for now, and the company said it is working on extending the marking to them.
This is not limited to the chat interface. It covers Claude, Claude Code, Claude Cowork, Claude Tag and the API, and it also covers the models when they are used through AWS, Google Cloud and Microsoft Foundry. Anthropic applied it worldwide, not only to users in the European Union.
What is actually embedded
In text it is a hidden statistical signal that, according to Anthropic, does not change the meaning, quality or readability of the response. In supported files such as images, signed metadata is added using C2PA, an open standard for recording where a piece of content came from. The model itself does not know its answer is being marked.
Anthropic said it will publish the technical details for detecting the mark, so the ability to check will not stay in its hands alone.
Why now
The driver is Article 50 of the European AI Act, which requires AI-generated content to be marked. Its transparency obligations took effect on 2 August 2026, the same date from which every new Claude model ships marked. Anthropic chose to apply the requirement to users everywhere rather than only to the European market.
What it changes for a small business
A business owner who uses Claude to write the marketing copy on a site, the product descriptions, a client proposal or a reply to an enquiry is now publishing text that carries a mark. Nothing about that is prohibited, and Israel imposes no duty to declare it. What changed is that the question of whether a text was written with AI stopped being a matter of instinct.
The sensitive case is a business that sells writing, translation, editing or design as the deliverable. A client contract that restricts AI use becomes checkable. It matters here that the mark shows Claude touched the text, not that it wrote the text from scratch. Someone who writes their own copy and then runs it through the model to proofread or translate can end up with the same result as someone who asked the model for the whole thing. Better to agree in writing with the client up front on what is allowed, rather than deal with a check after the fact.
Requests for quotes and tenders already ask about AI use. An accurate answer beats a comfortable one.
The limits of the mark
Anthropic was explicit that the marking is not conclusive. It survives copy and paste and may survive light editing, but a heavy rewrite or a translation can strip it, and very short passages will not be detected reliably. In files, a screenshot or a re-save in another format can remove the signature.
In the other direction, the absence of a mark says nothing about a human having written the text. Older models are unmarked, and so are models from other companies that have not adopted similar marking. The mark shows Claude was involved. It does not show who is responsible for the content, how much of it the model wrote, or what was edited afterwards.
What to do this week
Go through your client contracts and check whether any of them has a clause about AI. Decide what your policy is and write it in one line, even if it reads like "we use AI for drafts and editing, and a person reviews every deliverable". A client who asks gets a ready answer. And anyone handing work to a client who ruled out that kind of use should know the check is now available.