Skip to content
il Cantonale

Independent digital newspaper of Italian-speaking Switzerland

World Artificial intelligence transparency

Anthropic puts an invisible watermark in the text Claude writes

The mark is embedded in the words themselves and survives copy and paste. The measure answers article 50 of the European artificial intelligence regulation, applicable since 2 August, but the company is extending it worldwide.

by Gottardo 12 August 2026 3 min read

Anthropic has begun inserting a machine-readable mark into the text produced by its new models. It is not a caption, not a footnote and not visible to the eye: it is a statistical signal embedded in the choice and arrangement of words. Anyone reading the text will not notice it; a system that knows to look will find it.

The practical consequence is that the mark survives copying and pasting. That is the difference from metadata, which is stripped as soon as text moves from one document to another or is pasted into a form. Anthropic maintains that the marking does not alter the meaning, quality or readability of a response. For generated files in svg, png and jpg formats the company instead uses signed metadata following the open C2pa standard, adopted by part of the industry to certify the origin of images.

A European obligation, applied everywhere

The measure answers article 50 of the European artificial intelligence regulation, the transparency provision that became applicable on 2 August and requires providers of generative systems to make artificially produced content recognisable. Anthropic has chosen to apply the marking worldwide and not only to European users. It is a recurring decision among large platforms: maintaining two different behaviours depending on jurisdiction costs more than aligning everywhere with the strictest standard.

The company has said it will publish the technical detail needed to detect the mark, so that users and third parties can check a piece of content. It has also set out the limits, which are substantial: a text may carry no detectable mark if it came from an earlier model, if it was heavily rewritten, or if it is too short for the signal to register. It follows that the absence of a mark proves nothing. The system can confirm that a text came from Claude; it cannot rule out that it came from another artificial intelligence, nor establish that a person wrote it.

That is worth keeping in mind, because the most natural use of such a tool is precisely the one that does not work: running a school assignment or a job application through it to find out whether it is genuine. The marking makes traceable what comes from a provider that has decided to mark it. About everything else it says nothing.

Comments

No comments

There are no comments yet. Yours can be the first voice.

Leave a comment