A machine can hide a signature in its word choices: at each step it quietly prefers a pseudo-random ‘green’ half of the vocabulary. You can’t see it — but a counter can. Slide (or tap) the watermark strength and watch the green-token count climb until the z-score screams ‘machine.’
A faithful sketch of the Kirchenbauer et al. (2023) LLM text watermark. Each token position hashes the previous token to seed a split of the vocabulary into a green list (fraction γ=0.5) and a red list; a watermarked model biases toward green. Detection needs no key beyond γ: count green tokens s over T tokens and compute z = (s − γT) / √(Tγ(1−γ)). Human text sits near z≈0; watermarked text drives z past 4 (astronomically unlikely by chance). The strength slider raises the green-bias; a fail-loud self-check throws unless z≈0 at strength 0 and z>4 at full.
A fixed synthetic token stream with a deterministic hash-partition stands in for a real tokenizer/model; γ=0.5 and the bias curve are illustrative. The green/red partition, the count, and the z-statistic are computed exactly — the detection math is the real thing.