Domestic Use Cases for AI Verification

The topic of AI verification is currently well-studied within the framework of international agreements, but this presents a subtle kind of danger, where these international verification mechanisms have their value intrinsically tied to an optimistic environment of great-power diplomacy that states (especially frontier states like the USA and China) may or may not even proceed with, due to factors such as political differences, etc.

We, the researchers (Joaquin Lorenzo De Guzman and Tobias Martin), aim to create a paper that investigates the inverse situation: "What benefits do DOMESTIC verification mechanisms have for a country's OWN INTERESTS, independent of any international agreement?". This is a lesser-studied topic, but one of equal importance, nonetheless.

We aim to have our audience consist of policymakers and technical agencies in mid-sized states or "middle powers" (Japan, the Netherlands, Singapore, South Korea, Switzerland, the UAE, and their peers) and we aim to create this paper with the intention of helping guide politicians to decide what verification capabilities to build, pilot, or watch within their country's own borders. Any benefit to a future US-China agreement is treated as spillover.

We have come up with a tentative list of seven priority verification mechanisms selected from a longer list of twenty-four, each tied to the domestic problem it solves. We also present a quantitative scoring matrix for prioritizing significant mechanisms, and three primary country archetypes plus one supplementary archetype. These archetypes are the entry point to a country-agnostic decision framework, where we intend to describe any state's position in the AI supply chain, and to help determine which use cases apply to it. This can therefore enable these states to identify which mechanisms are worth its money.

The latest project proposal we made for our research sprint is linked here. We are still in the early stages of the project (3rd draft).