Real-world effectiveness of hardware-dependent verification methods

Determine the real-world effectiveness of hardware-dependent verification methods for international AI governance, such as chip location tracking and chip-based reporting, in detecting unauthorized AI training runs and unauthorized data centers under realistic operational conditions.

Background

The paper categorizes verification methods into national technical means, access-dependent methods, and hardware-dependent methods. The hardware-dependent category includes proposals like chip location tracking and chip-based reporting, which would embed governance functionality into AI-capable hardware.

While these approaches could, in principle, offer robust and privacy-preserving verification mechanisms, the authors note that unlike mature methods used in other domains, their actual performance and reliability in real-world AI governance contexts have not yet been empirically established. Establishing their effectiveness is crucial for policymakers considering reliance on such mechanisms.