The watermark is Anthropic's half. The labeling duty is yours.

Anthropic published how it watermarks Claude's text output on August 14. The coverage read it as a detection story: can a ghostwritten post be caught, will subscribers quit. The half of the rule that changes a decision this quarter belongs to 152 companies that are not AI labs, and it has been in force since August 2.

The mechanism, briefly

Claude now steers the randomness behind its word choices using a key. To a person the text reads normally, while carrying a signal Anthropic can test for. It covers Claude, the API, Claude Code and Claude Cowork; models released after August 2 ship with it and older ones are being retrofitted.

Only one half of the rule has a vendor

The watermark exists because of the EU Code of Practice on Transparency of AI-Generated Content, published July 31 and applicable from August 2. Around 190 organizations signed. Eighty-two signed Section 1 as providers: Anthropic, Google, Meta, Microsoft, Mistral and OpenAI. A hundred and fifty-two signed Section 2 as deployers, and that list runs to Lufthansa, Getty Images, Lenovo, Bulgari and Iberdrola. The Code itself is voluntary; the binding floor under it is the AI Act's Article 50 transparency rules, applicable from the same August 2 date to anyone selling into Europe, and signing early is how these companies chose to meet them.

Read the second list again. An airline, a stock photo library, a hardware manufacturer, a jeweler and a utility. The Code splits the duty: the lab marks the output, the company publishing it labels it. Anthropic has finished its half and documented exactly how. Your half has no vendor and no launch date.

For three years the working assumption inside marketing, sales and legal was that AI-drafted text was indistinguishable from the human kind, so disclosure was a matter of taste. That was a rational read at the time. No mechanism existed to mark the output, and volunteering the information was pure downside. The mechanism exists now, the clock has started, and the policy you never wrote is the one you are about to be measured against.

The objection is right and it doesn't help you

The watermark is weak, and Anthropic says so plainly. A full rewrite removes it. The signal thins on factual passages where there are fewer word choices to make. Small samples do not work. The detection API has not shipped; the company's own phrasing is that it is still working out the details. Nobody is running forensics on your press release next Tuesday.

The law underneath doesn't move, though. Article 50 was never conditional on someone catching you, and what changed is that “nobody can tell” stopped being a safe planning assumption. If your company signed the Code, the commitment started August 2, and if you sell into Europe, Article 50 reaches you on its own. Everyone else meets it secondhand, through the AI clauses that EU-exposed customers' procurement teams are drafting right now. The first time it costs you, it will arrive as a procurement questionnaire rather than a lawsuit.

What I'd do about it

The work is small and specific. Your general counsel and your CMO need one written AI disclosure policy, a record of which customer-facing deliverables were AI-drafted, and language covering it in the MSA. Do it this quarter. Reconstructing a record you never kept is the expensive version, and it lands on someone else's deadline.

The larger pattern is worth naming. Labs will keep shipping provenance tooling, because a regulator asked and it costs them almost nothing. Companies downstream will keep mistaking that for compliance. Provenance is quietly becoming a property of the content supply chain, and anyone who puts words in front of a customer is in that chain whether they signed anything or not. Where that duty sits internally is the same argument as who should own the AI portfolio, arriving through a side door. If your teams generate customer-facing content faster than your policy can describe it, let's talk.