Refined Detection for Gumbel Watermarking
2026-03-31 • Machine Learning
Machine LearningCryptography and Security
AI summaryⓘ
The authors introduce a new, simple way to detect watermarks in text generated using the Gumbel watermarking method by Aaronson (2022). They show that their detection method is nearly the best possible among all general detection methods that don’t depend on the specifics of the model. This conclusion holds under the assumption that the predicted next word distributions are independent and identically distributed (i.i.d.). Their work helps improve understanding of watermark detection in AI-generated text.
Gumbel watermarkingwatermark detectionmodel-agnosticnext-token distributioni.i.d. assumptionnear-optimalityAI text watermarkingAaronson 2022
Authors
Tor Lattimore
Abstract
We propose a simple detection mechanism for the Gumbel watermarking scheme proposed by Aaronson (2022). The new mechanism is proven to be near-optimal in a problem-dependent sense among all model-agnostic watermarking schemes under the assumption that the next-token distribution is sampled i.i.d.