Paper 2026/2291 Original: eprint.iacr.org/2026/2291 ↗

Evidence on Demand: Selectively Disclosable Watermarks for Generative AI

Abstract

Cryptographic watermarking for generative AI equips a generative model with a hidden mark that identifies its outputs as machine-generated, without affecting output quality. This study proposes a new cryptographic watermarking method with selective disclosability. Under this scheme, detection is restricted by default to the model owner, who can selectively issue a short token for any single output. This token enables any third party to verify that the specific output is watermarked, while guaranteeing that all other outputs remain undetectable, even to the token holder. This scheme is particularly useful in scenarios where a provider must certify a single controversial output without compromising the privacy of other outputs. In addition to two existing security properties, namely robustness and unforgeability, we formalize two security properties, undetectability of undisclosed data and disclosure soundness, to achieve our goal. We then propose our construction based on prior work (Lin, Shahabi, and Song, CRYPTO 2026) and prove it satisfies the four security notions. Our construction is broadly applicable to autoregressive generative AI models, including LLMs.