Text watermarking in LLMs relies on pseudo-randomly partitioning the vocabulary into 'green' and 'red' token sets based on the preceding n-gram context. During generation, the model's logits are subtly adjusted to favor green tokens without altering semantic meaning.
When evaluating a text sample, a detector computes the z-score of green token frequency against the expected uniform distribution. If the z-score exceeds safety thresholds, the text is positively identified as AI-generated.
Get Tech Pulse Daily in Your Inbox
Join 45,000+ engineers, founders, and tech leaders receiving high-signal daily breakdowns directly from major publishers.
Zero spam. Unsubscribe anytime in one click.
Anthropic's innovation lies in dynamic entropy control: in highly constrained generation tasks (such as code syntax or mathematical proofs), watermarking strength automatically dials back to prevent invalid token outputs.