Watch a 47-minute lecture from the Simons Institute where Noah Golowich from MIT presents groundbreaking research on watermarking language model outputs with provable guarantees. Explore innovative approaches to detecting AI-generated text through a watermarking scheme that achieves both undetectability and robustness to edits. Learn about the cryptographic concept of undetectability introduced by Christ, Gunn & Zamir, and discover how this new scheme handles adversarial insertions, substitutions, and deletions in watermarked text. Understand the technical implementation of indexing pseudorandom codes and their role in creating robust watermarks when working with large alphabets. Gain insights into the generic transformation process from codes to watermarking schemes for language models, all while using weaker computational assumptions than previous approaches. The presentation, part of the "Alignment, Trust, Watermarking, and Copyright Issues in LLMs" series, represents joint work with Ankur Moitra and offers a significant advancement in the field of AI text detection and watermarking.
Overview
Syllabus
Edit Distance Robust Watermarks: beyond substitution channels
Taught by
Simons Institute