Claude Now Signs Everything It Writes. Anthropic Is Watermarking AI Text, and the EU Made It Global.

Starting with new Claude models released on or after 2 August, Anthropic is embedding invisible, machine-readable watermarks into text generated by Claude, and attaching digitally signed provenance metadata to the files it produces, following the open C2PA standard already used for images. There is no opt-out, and it is global. The reason it reaches everyone is a European one: the EU AI Act's Article 50, which requires AI outputs to be marked as such.

What the mark actually is

This is not a visible label or a disclaimer. It is a statistical signature woven into the choice of words, invisible to a reader but detectable by a machine that knows to look. For files like PNG, JPG and SVG, Claude attaches signed metadata recording the content's origin. Crucially, the mark proves processing, not authorship: it tells you text passed through Claude, not whether a human wrote or edited alongside it. That is a narrower claim than the headlines imply, and an honest one.

What it catches, and what it does not

Watermarking of this kind is genuinely useful for the easy, high-volume case: bulk AI text moving through systems that check for the mark. It is also fragile at the edges. Paraphrase the text heavily, run it through another model, or translate it, and the signal degrades or disappears. So it is better understood as friction and provenance than as a lie detector. It raises the cost of passing machine text off as untouched, and gives compliant systems a way to label AI output, without ever being airtight.

Why one company’s policy is really a standard

The most important part is not the technology, it is that a European law is now setting a global default through a private company's product decision. Anthropic did not build separate EU and rest-of-world versions of Claude; it turned marking on everywhere. Expect the same from every major lab, because maintaining two versions costs more than complying once. Provenance, who or what produced a piece of content, is becoming a built-in property of AI output rather than an afterthought, the same direction as the broader push to test and govern frontier models before release.

What to do with it

For most businesses this is a quiet infrastructure change, not a fire drill, but three things are worth knowing. AI text you generate may now carry a detectable mark, which matters if provenance or disclosure is part of your compliance. Watermarks can be stripped, so do not treat their absence as proof a human wrote something. And verifying where a document, image or number came from is becoming a first-class requirement rather than a nice-to-have. The label is arriving. The habit of checking it is the part still being built.