See how statistical text watermarking marks word choices — and how a detector adds up the evidence.
Educational simulation — read this first. A real detector needs the model provider's secret key and its exact tokenizer. Nobody outside Anthropic has Claude's key, so this tool cannot tell you whether real text came from Claude. Instead it uses its own key to demonstrate the mechanism faithfully: the same green/red partition and the same z-score math the published methods (Kirchenbauer et al.; Google's SynthID-Text) use. Paste your own writing and you'll see roughly chance (~50% green). Use the demo buttons to watch what a watermarked vs. an unmarked passage looks like under the same detector.
When a model generates text it samples each next word from a probability distribution. A red/green watermark seeds a pseudo-random generator with a hash of the previous word(s) plus a secret key, and uses it to split the vocabulary into a green list (fraction γ) and a red list. During generation the model quietly adds a bias toward green words whenever the wording is a toss-up. The reader notices nothing, because there were many equally-good words and it just leaned toward some.
To detect, you re-derive the same green list at each position (you need the same key) and count how many words actually landed in it. Under no watermark, that count sits near γ·T by chance. Under a watermark, it sits well above. The z-score below measures how many standard deviations above chance the count is — a large z is the statistical fingerprint.
This tool derives the split from hash(previousWord + currentWord + key), which
is a faithful toy of the real "hash the context to seed the partition" step. It runs entirely
in your browser; nothing is sent anywhere.