AI Writing Now Comes With a Hidden Watermark

A guy in one of my workshops this spring asked me point blank if anybody could tell he was using AI to write his listing descriptions. The honest answer at the time was sort of. The detector tools out there were mostly guessing, and they were wrong often enough that you couldn't hang anything on them.

That's starting to change. Anthropic, the company behind Claude, turned on an invisible watermark in the text its newer models write. It kicked in August 2nd and they're doing it everywhere, not just in Europe. The mark doesn't change how the writing reads. You can't see it. It sticks around when you copy and paste, and it survives light editing.

The reason is a European law. Article 50 of the EU AI Act took effect August 2nd and it requires AI generated content to carry a machine readable mark. Anthropic signed onto the code of practice that goes with it, and instead of running one version of Claude for Europe and another for everybody else, they just turned it on for the whole world. Older models have until December 2nd to catch up. OpenAI and Google are sitting under the same rule, so I'd expect the same thing from them before the year is out.

What this doesn't mean

Nobody can paste your website into a box right now and get a verdict. There's no public scanner sitting out there flagging your About page. And the mark itself only says a Claude model touched that text. It says nothing about whether you wrote the first draft, whether you rewrote every line of it, or whether the ideas were yours.

So if you've been using AI to help you write, you're not in trouble. I use it every single day and I'm not going to stop.

What it does mean

The guessing part is ending. Maybe not this month, but the whole industry is moving toward text that carries its own receipt, and once that's normal, the question changes. It stops being can anyone tell and becomes do I care that they can.

For most of what a small business puts out, the answer is no. Nobody's offended that your website copy got help. Your customers assume it already.

Where I'd actually pay attention is the personal stuff. A proposal you're sending to a client who's about to hand you thirty grand. A note to a customer whose situation went sideways. An apology. Those are the ones where getting caught outsourcing it costs you something real, and honestly, those are the ones you should be writing yourself anyway.

The other place to look is contracts and applications. Some RFPs and grant applications have started adding language about AI generated submissions, and a few of them ban it outright. If you're bidding on anything with a public entity or a big institution, read that section before you paste anything in. That clause wasn't in there two years ago.

One more thing. Don't go hunting for a tool that strips the watermark out. In the EU that's flatly prohibited, and while I doubt anyone's coming after a landscaping company in Sylvania over it, if you're going to that much trouble to hide something, that's usually a sign the thing you're hiding is the actual problem.

The part people keep missing

A European law changed what an American company does for American customers, and nobody voted on it here. We've watched this movie before with the cookie banners on every website you visit. That was GDPR, also European, and it reshaped the whole internet for everybody.

So when somebody tells you the AI regulations don't apply to a small shop in Northwest Ohio, I'd take that with a grain of salt. They don't apply to you directly. They apply to the tools you use, and that ends up being the same thing.