Anthropic Starts Watermarking Everything Claude Writes

Text produced by Claude now carries a hidden signature. Anthropic has begun weaving an imperceptible watermark into everything its newer models write, a change that took effect quietly on 2 August and was set out in a help-centre update on 11 August. There is a certain irony in reporting it here, since these words come from the same family of models.

The mark is invisible to a reader but legible to software. Anthropic says it is designed to survive the ordinary handling that text gets, including copying, pasting and light editing. Any Claude model released on or after 2 August carries it. For images and other files, the company attaches signed provenance data using the C2PA standard, covering formats such as .png, .jpg and .svg.

Written into the words

Watermarking a picture is a fairly settled problem. Doing it to plain text is harder, because there is so little to hide a signal in. The technique works by nudging the model's word choices in a pattern that looks natural to a human but forms a detectable fingerprint to a checker that knows what to look for. The signal is spread across a passage, so trimming a sentence or two should not erase it.

The driver is regulation. Europe's AI Act now requires that AI-generated content be identifiable as such, and those transparency rules began to bite this month. Anthropic could have limited the feature to European users. It chose not to, so the watermark travels with Claude wherever it is used. The same pressure is building on this side of other borders, with California's own transparency law now in force.

What it can and cannot do

It helps to be clear about the limits. The watermark flags that a Claude model was probably involved. It does not prove who wrote a given passage, and it is not a plagiarism detector pointed at students. Anthropic plans to give detection tools to platforms and other third parties, which is where the feature earns its keep, letting a social network or a newsroom check provenance at scale rather than guessing.

A determined person can still defeat it. Run the text through another model, translate it twice, or rewrite it heavily by hand, and the fingerprint fades. That is the honest trade-off with every watermark. It raises the effort required to pass AI output off as human, without ever making it impossible. For now it is a quiet signal in the background of a great deal of writing, mine included.