Anthropic to Add Statistical Watermark to Claude Text

Anthropic has announced that future versions of its Claude models will embed a statistical watermark in generated text, a change the company says is driven by the EU AI Act. The watermarking will be applied globally at launch, because Anthropic does not yet have a durable way to scope it by region.

The mechanism exploits the fact that large language models choose words partly at random. When several words are equally plausible continuations—"overcast" versus "grey" after "The weather today was cold and"—the model picks one using a source of randomness. The watermark replaces that arbitrary source with one derived from a secret key and preceding words. The output remains random in practice, but it now carries a pattern that anyone holding the key can check to estimate the likelihood Claude generated the text.

The technique is a version of Google DeepMind's SynthID-Text approach, published in Nature in 2024, and belongs to a family dating to a 2022 proposal by Scott Aaronson. Anthropic cites internal testing and DeepMind's own Gemini traffic experiments as evidence that watermarked text is indistinguishable from unwatermarked output in quality, creativity, and readability. Nothing is added to the text—no hidden characters or extra tokens—so there is no added cost or latency. The watermark cannot be traced to a specific user, organization, or chat.

The method has clear limits. Detection is unreliable on short passages, where there are too few random choices to establish a pattern. It is sparser on factual text—there is only one correct continuation to "Isaac Newton's most famous work was called Principia"—and in code, where exactness dominates. In proofreading tasks, where most words are the human author's, the watermark may have too little to attach to. A complete rewrite removes it; light editing probably does not. A watermark can indicate Claude was likely involved, but cannot distinguish "Claude wrote this" from "Claude heavily edited this," nor confirm whether text was human-written.

Anthropic describes the watermark as a compliance measure. The EU began requiring AI providers serving its market to mark AI-generated content on August 2, and Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026, alongside roughly 190 other signatories. Older Claude models fall under a transition period and will receive watermarking in coming months.

A watermark detection API is forthcoming, with implementation details still being worked out. Image and other file outputs will carry C2PA content credentials—cryptographically signed metadata notes rather than embedded watermarks. Anthropic contrasts its key-based watermark with third-party AI detection software, which relies on stylistic tells like overused phrases and has no access to the watermark key.

One open question: how the detection API will be exposed and who gets access. Anthropic says it is still determining the details.