Where an AI watermark actually hides.

Between the letters you can read, text can carry characters that display as nothing at all: invisible spaces, joiners, direction marks. They hold real positions in your text, and they travel with it through every copy and paste.

Many AI tools emit them, embedded in what you paste without you ever seeing it, and anyone who knows to look can check for them in seconds. They are the easiest tell to catch, and the easiest to remove completely.

Every file carries a wrapper that describes it: what made it, when, with what. When an AI tool generates a file, it writes a record into that wrapper naming itself. Claude does it, and so do OpenAI, Google, Adobe and the rest, most of them behind one shared industry format called C2PA.

We strip that record, along with EXIF, XMP and generator tags, then read the file back to confirm nothing is left. Your picture or document comes out byte for byte identical, and on a file Claude made, this is Claude’s mark coming off, with proof.

This is the mark Claude, and others, apply to text itself: the one in the news, and the one nobody can point at, because nothing is inserted at any point. Here is the trick. As a model writes, it constantly reaches words where several options would read equally well: the sky was grey, overcast, gloomy. A normal model just picks one. A watermarking model hands that pick to a secret key, every time. Each choice reads naturally on its own, but across a few hundred words the picks line up into a hidden pattern, and whoever holds the key can test any text for it.

That pattern lives only in unbroken stretches of the original words, which is why you cannot beat it by swapping synonyms or asking a chatbot to reword: a casual rewrite leaves long stretches of the original untouched, and every one of them still carries the signature. Beating it takes an engine built for this one job, one that dismantles the word sequences the mark rides on. That is what we built, and its rules are below.