What happened
On 11 August Anthropic updated its documentation and confirmed what Europe had until now been discussing in the future tense: Claude marks the text it generates. Not with a label in the interface, and not with a note at the end of the response. The watermark is woven into the text itself, invisible when you read it, and it travels with the content when you copy it. To generated files — .svg, .png, .jpg — the company adds signed provenance metadata in the C2PA standard.
Two things in that announcement matter more than the fact itself.
The first is reach. The documentation states that marking will cover the output of supported models everywhere Claude is available, worldwide. There is no separate compliant version of the product for Europe and an unmarked one for everyone else. Models placed on the EU market on or after 2 August 2026 carry the marking from the start; for the earlier ones, Anthropic is adding it now.
The second is that this is not one provider's move. The Code of Practice on Transparency of AI-generated Content — the instrument the Commission and the AI Board deemed adequate for demonstrating compliance with Article 50(2), (4) and (5) — had been signed by around 190 organizations by the end of July. In the section written for providers, the Commission lists, alongside Anthropic, Google, Meta, Microsoft, OpenAI, Mistral and Cohere among others.
The date itself is not new. A week ago we took Article 50 apart paragraph by paragraph: the transparency obligations apply from 2 August 2026, and the machine-readable marking in paragraph 2 was given a transition until 2 December — but only for systems placed on the market before August. What is new is that the provision has stopped being a calendar entry and has entered the material your company sends out.
Our read
The mark that has just appeared in your text is evidence — only not yours.
It is asymmetric by construction, and that is a description, not an accusation. Anthropic itself writes that a detected mark does not confirm the full provenance of the content: it says the material passed through Claude, not that Claude wrote it. In the other direction it works even less well. The mark stops being detectable once the text is heavily rewritten, paraphrased, translated or mixed into other writing, and file metadata is lost when the format is converted. Which means it disappears in exactly the cases where a human did the most work on the text.
On top of that, you cannot read it. The same documentation states that the company is only now working on making detection available to users and third parties, and that the details will land in future technical documentation. So as of this week you have a signal in your own material that you cannot check yourself.
Put the two together. The presence of the mark can raise a question about your document. Its absence will not answer that question in your favour, because it just as easily means someone simply edited the text. A tool that can only open a doubt and never close it is not fit to be your internal control. If your AI usage policy says "we verify the markings", what you have written down is a control that works in one direction — and not the one you need it in.
What has not changed in the process: your own obligations. The marking duty in paragraph 2 is written for the model provider, and the provider has just discharged it. As a deployer you still have paragraph 3 and paragraph 4 — emotion recognition and biometric categorisation, and deepfakes and model-generated text published in order to inform the public on matters of public interest. A client report and a commercial proposal fit into neither, and nothing this week moved that.
Something else does change, and here we label this plainly as our expectation, not a finding: what changes is what the other side can infer. Once detection becomes widely available, a client, a competitor or a reviewer gets a cheap way to raise a question about your material. The law requires you to disclose nothing in that situation. A contract in which you promised authorship, or a tender clause about preparing the bid on your own, may require something else — and that is a commercial conversation, not a regulatory one, and it happens with no authority in the room.
The practical conclusion is short: the evidence that works in your direction will not be the provider's mark. It will be the record you keep yourself — what was produced with a model, who approved it and when.
Why it matters
Private Equity
For a fund this is not a risk line but a question about process hygiene at companies that sell intellectual work: agencies, software houses, consultancies, design studios. If a portfolio company delivers documents to clients that were produced with a model and keeps no record of it, that is not a legal problem — it is an unrecognized contractual exposure that surfaces at the first complaint or at the renewal of a master agreement. The question for the next portfolio review fits in one sentence: on the product being sold, does anyone write down where a model was used. The answer "nobody tracks it" is as useful as "it's in the register", because both tell you how much work is left.
Enterprise
A large organization has an AI usage policy, and most often it is written around the question "are we allowed". This week adds a second, harder one: what do we show when someone asks about a specific document from six months ago. That question is settled not in the policy but in whether the document workflow has anywhere to put such a note at all.
It is also worth reviewing client contracts for authorship clauses and for confidentiality of input material. A clause written three years ago may have promised exclusively human execution in a way nobody read literally at the time, because there was no way to check it. Now there is starting to be. This is how we set up AI governance in a large organization: first a register of where the model actually touches the product, then the contract wording, and the policy last — because the policy is meant to describe reality, not the other way round.
SMB / mid-market
This is where it is simplest and cheapest, provided you do it now rather than after the question arrives. You need neither a tool nor a lawyer. You need one column where you already run your business — in the CRM, in the proposals sheet, in the project folder — answering whether a model worked on this material and who read it before it went out. A company that can show that is having an ordinary conversation about quality. A company that cannot is having a conversation about trust, and that one is far more expensive.
One move this week
Take the last three documents that went from you to a client and were produced with a model — a proposal, a report, a specification. Answer two questions. Is it written down anywhere that a model worked on them, and is it known who approved them before they went out. If the answer to both is "no", this is one afternoon of work: a column, a date, a name. Reconstructing it after the fact in the week the client asks costs many times more and looks exactly like what it is. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI.