Text Watermark Detection Does Not Require Rerunning the LLM

rasbt · x · 2026-08-18

Clarifies the technical mechanism behind text watermarking detection, correcting the misconception that it requires rerunning the LLM. Based on methods like SynthID-Text, detection only requires reapplying cheap 'watermarking functions' that score specific terms. If the text exceeds a score threshold of high-scoring watermark words, it is likely watermarked.

Original post →

More from Infra

Infra channel →