Anthropic Now Watermarks Every Claude Output, Globally

Can you tell whether a paragraph came out of a language model? Anthropic has decided that question shouldn't depend on your intuition. Every Claude model published after August 2, 2026 now embeds an invisible marker into the text it produces, and the policy applies worldwide rather than only in Europe.

The scope covers the whole product line: the Claude platform API, the Claude app, Claude Code, Claude Cowork and Claude Tag. The company laid it out in a support document on August 10, and the news spread the following day.

Two different technologies, two different weaknesses

For text, the watermark lives inside the writing itself. Anthropic's phrasing is that it "travels with the text when copied and pasted, and may survive some editing." For files such as PNG, JPG and SVG, the company attaches a digital signature based on the C2PA standard instead.

Those are not equally durable. C2PA data sits in file metadata and disappears the moment someone takes a screenshot. A text watermark is a statistical pattern in word choice, invisible to a reader but recoverable through analysis. Neither survives a determined attempt to strip it.

The EU wrote this rule, everyone gets it

Anthropic frames the move as following through on the EU AI Act transparency code it signed. The relevant obligations took effect on August 2 and require machine-readable marking of AI-generated content.

What's notable is that the company chose not to run one Claude for Europe and another for everyone else. That's the Brussels effect in miniature: a regional rule becomes a global default because maintaining two behaviours costs more than complying everywhere. Competitors are scattered on this. Google DeepMind open-sourced SynthID; OpenAI has claimed a detector with 99.9% accuracy for years without shipping it.

A marker is not proof

Anthropic's own caveat is the most useful part of the announcement. A watermark does not establish that a text was written by AI. Someone may have written it themselves and asked Claude to fix the grammar, translate it, or tighten it. The reverse also holds: heavy editing, reformatting or translation into another language can wipe the signal out.

The verification tooling isn't out yet either. Anthropic says it is building tools for users and third parties, but has published nothing. Right now the half of the system that marks content works, and the half that reads it does not.

What this means if your team publishes with AI

Nothing here demands a change of course, but marketing, legal and education teams should think it through once. If you sell into the EU, deliver content to European clients, or answer procurement questionnaires that ask about AI usage, the ground has shifted slightly under you.

The risk we'd actually worry about isn't the watermark. It's the person misreading it. A client, a teacher or a reviewer runs some future tool, sees "Claude signature detected," and treats that as an accusation. The marker will be present in dozens of perfectly ordinary uses, from spellchecking to translation, and a detection result carries none of that context.

The practical answer is to stop treating AI usage as something to hide. Write down which tasks use a model, who reviews the output, and who signs off on accuracy. Teams that can answer those three questions in a sentence have nothing to fear from a watermark. Teams that can't had a problem before this announcement too.

Sources: TechCrunch, The Decoder, Anthropic support documentation