Ordinary words… carrying a hidden mark
What changed in Claude? And how should we read the watermark, without the hype?
01How does it work?
An illustration, not a testThe model chooses between suitable words. The choices repeat in a pattern that a special detector can recognize.
The day was cold and cloudy.
SynthID-Text leaves a signal spread across the text.
02Who uses it for text?
Based on official statementsnot every older version.
we do not generalize to every interface.
images and audio.
this does not prove SynthID use.
this does not prove SynthID use.
Not announced ≠ not present. The hosted service may behave differently from the same model run locally.
03How does it affect you?
It can reveal that AI took part in writing or editing.
In Claude: it does not reveal who you are, and on its own it does not prove you did not write the text.
04How should you handle it?
Use the ideas, then write and check the facts. Heavy changes can weaken the signal.
An open model running without a mark can reword the text; the result still needs review.
Cleaning up spaces is not enough. And an add-on does not guarantee getting past Claude's detector.
Rewording can make the mark harder to detect, but it is no guarantee. And a missing mark does not prove a human wrote the text.