Claude’s text now carries a mark you cannot see

On August 14 Anthropic explained Claude’s watermark. Since August 2 the European Union has required anyone selling AI on the market to mark what the model writes. The company signed the code of conduct with about 190 names, and because it still cannot turn the mark on for Europe only, it turned it on worldwide.

How it works, in plain English. The model picks the next word. Sometimes both will do: “cloudy” or “overcast.” That choice, which a reader never notices, leaves a pattern. With the key, you can score the odds that Claude wrote it. Without the key, the text is just text. The company says quality does not change, the price does not go up, and it does not identify the person or the chat. The method is a version of Google DeepMind’s SynthID-Text, published in Nature in 2024.

The other look is in Anthropic’s own write-up, and it is worth reading out loud. The mark does not say “Claude wrote this.” It says “Claude probably had a hand in this.” A short note barely takes a mark: few choices, little signal. Tight factual prose — “Newton’s work was called Principia…” — leaves no room to choose. Code, even less. If someone only asks for a comma fix, the mark barely lands. If they rewrite it word by word, it disappears. A translation by Claude, yes: every word is its pick.

Detection is not out yet. The company promises an API. Images get something else: a C2PA stamp on the file, the same standard cameras and photo apps use. You can read it. It is not hidden in the pixels.

If you just want the news: Claude’s text now carries an invisible mark, because of Europe, everywhere. It does not name the author. It names the model — and it vanishes if you actually rewrite.