Contract monitoring: governing AI via separation of powers

AI systems can be governed through "contracts" that specify what actions they're allowed to take, similar to how legal contracts bind humans. This paper demonstrates that such contracts can be enforced through a system of checks and balances—assigning different monitoring and oversight roles to different agents rather than relying on a single control mechanism.

Separation of powers improves AI safety: By dividing responsibilities between "monitor agents" (who watch for violations) and "judges" (who make final decisions), the framework reduces the risk that any single component could be compromised or fail. This mirrors how democratic institutions use checks and balances to prevent abuse of power.

The framework provides measurable safety guarantees, allowing organizations to quantify how reliably an AI system stays within its permitted boundaries. Rather than making vague claims about safety, this approach enables empirical testing and verification of whether contracts are actually being enforced.

Computational resource allocation matters for safety: The paper shows that giving monitors and judges more computational resources than the worker agents they oversee helps prevent AI systems from circumventing oversight. This suggests that safety infrastructure requires adequate investment to be effective.

This approach works across diverse AI safety challenges, from preventing code security breaches to stopping AI systems from "escaping the box" (circumventing their constraints), indicating the framework's broad applicability to real-world AI governance problems.