New Claude models now watermark their text invisibly, worldwide

AI watermarking just became the default. Claude now marks the text it writes, invisibly, everywhere in the world.

How the marking works

Starting August 2, 2026, new Claude models embed a machine-readable watermark directly into generated text. A reader cannot see it. Anthropic says the mark stays in the text when it is copied and pasted, and may survive some editing. The technique is generally understood to work by subtly biasing which tokens the model picks during generation, in a pattern a detector can recognize later, though Anthropic has not published the specifics.

Files get equivalent treatment through a different route. Supported formats including PNG, JPG and SVG carry signed provenance metadata using C2PA, the Coalition for Content Provenance and Authenticity standard.

Why now, and why everywhere

The trigger is Article 50 of the EU AI Act, whose transparency code of practice became enforceable on August 2, 2026. What is notable is the scope: Anthropic is applying the marking to Claude products worldwide, not only to users in Europe. Rather than maintaining a marked European product and an unmarked one for everyone else, the compliance behavior becomes the global default.

From vibes to signal

Informal AI detection has so far run on style: readers treating surface features of prose, down to individual punctuation marks like the em dash, as evidence of machine authorship. That is guessing. A watermark is categorically different: a signal deliberately placed in the text at generation time, and it does not care how anyone writes.

The limits matter as much as the capability. Existing documents are not retroactively marked, and a watermark is not the same thing as a reliable detector in every situation; Anthropic hedges on how much editing the mark survives. Nothing here settles arguments about unmarked text from other models.

What it does change is the default for one widely used model family. If Claude output is pasted into something that will be checked, the check no longer has to rely on vibes. For schools, employers and publishers that have been improvising detection policy, that is the first piece of solid ground in a while.