Difficulty of Monitoring AI Inference for Dangerous Use

Characterize the practical difficulty and requirements for monitoring AI inference workloads to ensure they do not contribute to dangerous AI development or misuse, including the ability to detect prohibited research or capability accumulation in deployed models under an international ASI-prevention regime.

Background

The agreement’s verification approach focuses on preventing large-scale training, but the authors note that powerful existing models might still advance dangerous activities through inference. Monitoring inference may therefore be necessary to prevent prohibited progress or misuse.

The authors explicitly state uncertainty about how difficult inference monitoring would be, underscoring the need to assess detection capabilities, operational costs, and robustness against evasion.