Contract monitoring: governing AI via separation of powers

Enric Boix-Adsera

We propose an AI safety framework that binds worker agents to contracts specifying their permitted actions. We show how these contracts can be enforced and specified by assigning d…