A new arXiv preprint introduces VOW (Verifiable and Oblivious Watermark Detection), a method aimed at making LLM watermark detection more auditable. The authors frame the problem around first-party provenance audit (FPA), where an organization needs to examine sensitive text for provenance while keeping the watermarking process secure.
With symmetric-key watermarks, the provider holding the key is the only party that can run detection. That arrangement is awkward for regulated settings where generated text must be marked and detectable but auditors need assurance. VOW is designed to make detection both verifiable and oblivious, so an auditor can check a watermark without gaining the provider's full detection capability.
The abstract is brief and does not spell out the cryptographic construction or evaluation details. Still, the motivation is clear: provenance requirements are moving from a single provider's claim to a checkable, independent process.