AI watermarks now sit inside the words your firm produces with Claude. Anthropic said on 11 August 2026 that it marks the text and the files its supported models generate, across its own applications and through the cloud services run by Amazon, Google and Microsoft. It applies the marking worldwide rather than only in Europe. The mark is invisible, it leaves the meaning and the readability alone, and it travels with the paragraph when a fee earner copies it into a letter, a statement or a skeleton argument.
Behind the change sits Article 50 of the EU AI Act, whose transparency duties became applicable on 2 August 2026 and which we set out in what the EU AI Act now asks of your business. Models launched in the European Union from that date carry the marking when they arrive, and Anthropic says it is adding the same support to older ones. Images take a different route, using signed provenance metadata under the C2PA standard.
What a mark proves, and what it does not
Anthropic is careful about this, and you should borrow its care. A detected mark tells you the content may have been processed by Claude, which the company describes as not fully conclusive. It does not tell you who wrote the original, because the model may have done no more than tidy a paragraph a solicitor had already drafted. It does not tell you the text has stayed as it left the model either.
The reverse trips people up more often. Clean text proves nothing at all. An older model, a heavy edit, a passage too short to carry a signal, a platform that strips metadata on the way through, any one of those leaves a document with no mark in it and an AI tool somewhere in its history. A firm that treats a detector as a test of whether AI touched a file will reach the wrong answer in both directions, and so will an opponent running one against your bundle.
What this asks of your firm
The real change is that your use of these tools has become visible in a way it was not a fortnight ago. Anthropic says it is working towards letting users and third parties read the marks, so work drafted with Claude and sent to a client, an opponent or a court will carry something legible once those tools exist. That prospect is uncomfortable only for a firm that cannot describe its own process.
Your duty was never to keep AI out of the work. It was to check what came out of the tool and to stay answerable for the advice. The judiciary's guidance and the SRA's standards put that on the person who signs the document, and a mark buried in the text says nothing about whether the citations are real. It records how a draft was made, not whether it is any good.
So write down where these tools sit in the work. Name them, name the tasks they are allowed near, and record who read the output before it left the building. Tell your fee earners that the free account somebody opened on a personal login sits outside that record, and that it now leaves a trace in the client's document.
Then read what you have told your clients. A firm whose retainer or website promises that a human does all the drafting has a gap between the promise and the practice, and the mark is what exposes it. One sentence in your client care wording, saying where AI touches the work and who stays responsible for it, answers the question before a client thinks to ask.
The detection tools are not here yet. The marks are, in everything the model has written for you since the start of the month.
Anthropic sets out what it marks, how the text watermark behaves and what a detection does and does not establish in its own free-to-read note on how Claude marks AI-generated content.
If nobody at your firm can say which tools touched which files, that is the work we do first with practices your size: talk it through with us.