AI Watermarking: A Statistical Fig Leaf That a Local Paraphraser Can Strip in Seconds

Inside the Token-Level Bias That Turns AI Output Into a Traceable Signal

The explainer at declaude.org walks through the mechanics of AI text watermarking, a technique where a language model subtly biases its token-selection process according to a secret cryptographic key, producing a statistically detectable pattern that a verifier can later measure. This is timely because regulatory pressure—most visibly from the EU AI Act's transparency provisions and various content-provenance initiatives—has pushed major model providers to demonstrate some form of machine-detectable labeling for synthetic content. Technically, the approach relies on partitioning the vocabulary into 'green' and 'red' token sets per context and nudging generation toward the green list, so that watermarked text shows a measurable green-token surplus without a naïve human reader noticing. The article also touches on the reverse-engineering question—how many tokens of output an adversary would need to observe to infer or attack the key. Crucially, it frames watermarking as a probabilistic signal rather than an indelible fingerprint, which is precisely where the practical debate begins.