Preprint tests AI text watermark that withstands deletion attacks

Researchers report that a watermark for language-model text outperformed or matched baseline methods under several deletion and rewriting attacks in experiments with OPT-1.3B and Llama-3.1-8B-Instruct. The preprint tested the method, called RMCW, on the C4 and ELI5 datasets and reports strong detection on unmodified text. Deleting words can disrupt the positions a watermark detector relies on. RMCW embeds a mathematical pattern during generation through a secret-keyed partition of the vocabulary. Its detector searches surviving text fragments for local pieces of that pattern instead of reconstructing the entire original pattern. The reported results cover two models and two datasets.