Anthropic has started marking text produced by its Claude models so it can be identified as AI-generated later, and it plans to offer a way for outside parties to run that check. The company says it is doing this to comply with the European Union's AI Act, and it is applying the marking everywhere because it does not yet have a reliable way to limit it by region.
The marking is not a visible tag or a line of code stuck on the end. It works by nudging the model's word choices. When the model is deciding between two words that would both work (Anthropic's own example is "joyful" versus "cheerful"), those choices are steered in a pattern that a detector can later recognise. Anthropic says this does not degrade the writing.
So what does it mean if you have used Claude to draft a service page, a blog post or product descriptions? Practically, not much. This is not a Google penalty, it is not a legal problem, and nobody is coming to audit your About page. It is a transparency measure aimed at the AI companies themselves, driven by European law. OpenAI has indicated on a support page that it intends to do the same.
Two things are worth knowing. First, Anthropic says the marking has limits: it holds up less well on short fact-heavy text, on code, and on anything that has been heavily rewritten. If a human has edited your draft properly, which is what should be happening anyway, the signal weakens. Second, Forbes has reported a rush of "watermark removal" apps, and warns that scams and false claims are already mixed in with them. Do not go anywhere near those. Nothing about your website requires you to strip a watermark, and handing your copy to an unknown tool to do it is a bad trade.
The useful takeaway is the one that was already true: AI drafts should be edited by a person who knows the business before they go live. That improves the copy, and it happens to reduce how identifiable the text is as machine-written.
What to do: Nothing urgent. If you or a contractor produce site content with AI, keep a human editing pass in the process, and avoid any tool advertising watermark removal.
Reported by:
