Technical Deep Dive: How Anthropic's Claude Watermarking Algorithm Operates and Is Scanned

Pseudorandom Token Splitting and Green-List Probability Bias

Anthropic's implementation relies on a cryptographic hash of previously generated n-grams to split the vocabulary into 'green' and 'red' token candidate lists. During generation, the model soft-biases logit probabilities toward green-list tokens, embedding an invisible statistical signature into output text.

Because the bias key is shared securely with enterprise verification partners and compliance tools, scanners can compute a z-score across long documents to prove origin beyond a p-value threshold of 10^-6, even when users modify minor punctuation or phrasing.

Robustness Against Paraphrasing, Quantization, and Code Execution

Security researchers note that while prose watermarking is robust against light editing, applying similar watermarks to strictly deterministic code syntaxes presents unique challenges without risking syntax errors or performance regressions.