Short AI-generated texts like product reviews are hard to trace because existing watermarking for identifiers needs long texts or slow decoding.
The method maps each identifier to a longer code word from a secret codebook. It embeds the code word by adjusting token probabilities during generation. Extraction compares recovered bits to code words using a distance metric like Levenshtein, which allows deletions.
This recovers identifiers of 16 to 24 bits nearly perfectly and works well even on very short texts. The Levenshtein variant handles deletions and light rewrites. This makes it useful for attributing short deceptive AI text.