Skip to content

Anthropic Will Watermark Every Text From Claude: the Mark Survives Copying, but Nobody Defines "Some Editing"

1 min read
Share
Anthropic Will Watermark Every Text From Claude: the Mark Survives Copying, but Nobody Defines "Some Editing"

Since 2 August the Transparency Code of the AI Act has applied in the European Union, and it demands one seemingly simple thing: everything produced by a machine must carry a mark another system can read. Anthropic has now confirmed it will watermark the text its models generate, including Claude. Not through a press release, but through a quietly updated support page.

The mechanics are more interesting than the news itself. The watermark is not a sticker on the document - it is embedded in the text itself. The company writes that "because the watermark is part of the text, it will travel with it when copied and pasted elsewhere, and can survive some editing". For files, the open C2PA standard is used. The mark is applied at model level, meaning it is present whether the text came out of Claude, Claude Code, Claude Cowork or through the developer platform.

"Some editing" is the phrase worth stopping at. How much exactly do you have to rewrite for the mark to break? That the company does not say. All models released after 2 August carry the technology automatically, and support will extend to older ones - but the line between "signed text" and "unsigned text" stays undefined precisely where it should be clearest.

Anthropic is not alone. Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have also committed to honouring the European code. Last week the music platform Suno announced it would label songs made on it, after a string of legal challenges. Before that Substack teamed up with Pangram to flag machine-written texts - their chief executive Chris Best even coined a word for the phenomenon.

Notice the order. The labelling did not arrive when the technology was ready, but when lawsuits, unhappy users and a regulator with a deadline appeared. That is not an accusation - it is a description of how the industry works every time the cost of staying quiet exceeds the cost of admitting.

For a reader in Skopje, Zagreb or Belgrade this is more than news from Brussels. The European AI Act is exactly the type of regulation that reaches us with a few years' delay and is then translated into domestic legislation as a finished text. Companies are already building the mark in globally, because that is cheaper than maintaining two versions of the same model - so the benefit of a European rule reaches even those with no influence over its writing.

The question that remains is not technical. If the watermark survives copying but not rewriting, then it catches precisely the person who was not trying to hide anyway - and not the one the rule was written for.