Anthropic explains the Claude text watermark — and where it stops working

Anthropic published the mechanics of Claude’s text watermark on 14 August 2026. Rather than inserting characters, it replaces the randomness behind low-stakes word choices with a function of a secret key and the preceding words, leaving a pattern invisible to a reader and detectable to anyone holding the key. Anthropic identifies it as a version of Google DeepMind’s SynthID-Text, published in a peer-reviewed 2024 Nature paper, and says it carries no identifying information, needs no extra tokens and has no practical effect on quality, cost or speed. The limits are stated plainly: detection performs poorly on short passages, thins out on factual text where accuracy forces one right answer, and thins out on code, which has to be exact to run.

What it means A negative watermark result on a short answer, a factual paragraph or a code block means almost nothing — so any grading, moderation or disclosure policy built on "we checked and it came back clean" needs the passage-length caveat written into it.