[Submitted on 25 May 2023]

Download PDF

Abstract: Recent advances in the capabilities of large language models such as GPT-4
have spurred increasing concern about our ability to detect AI-generated text.
Prior works have suggested methods of embedding watermarks in model outputs, by
noticeably altering the output distribution. We ask: Is it possible to
introduce a watermark without incurring any detectable change to the output
distribution?


To this end we introduce a cryptographically-inspired notion of undetectable
watermarks for language models. That is, watermarks can be detected only with
the knowledge of a secret key; without the secret key, it is computationally
intractable to distinguish watermarked outputs from those of the original
model. In particular, it is impossible for a user to observe any degradation in
the quality of the text. Crucially, watermarks should remain undetectable even
when the user is allowed to adaptively query the model with arbitrarily chosen
prompts. We construct undetectable watermarks based on the existence of one-way
functions, a standard assumption in cryptography.

Submission history

From: Or Zamir [view email]



[v1]

Thu, 25 May 2023 02:57:16 UTC (46 KB)

Read More