Paper Image

Uncovering the Hidden Signals: A Popular Science Guide to Language Model Watermarking

Published on:

24 January 2023

Primary Category:

Machine Learning

Paper Authors:

John Kirchenbauer,

Jonas Geiping,

Yuxin Wen,

Jonathan Katz,

Ian Miers,

Tom Goldstein

Bullets

Key Details

Watermarks generated text while allowing high-quality output.

Detection algorithm is efficient, public, and statistically rigorous.

Watermark adapts to sentence entropy, avoiding low-entropy sequences.

Difficult to remove watermark without major edits to text.

Provides transparency while protecting language model IP.

AI generated summary

Uncovering the Hidden Signals: A Popular Science Guide to Language Model Watermarking

This paper proposes an efficient method to watermark text generated by large language models. The watermark makes synthetic text detectable from short spans of tokens, while false-positives on human text are statistically improbable. The detection algorithm can be open-sourced without the model itself. The watermark adapts to text entropy, weakly marking unpredictable text to minimize impact on quality.

Answers from this paper

Comments

No comments yet, be the first to start the conversation...

Sign up to comment on this paper

Sign Up