Who Watches the AI?

Autonomous AI systems, particularly those capable of recursive self-improvement, pose a significant risk of acting against human intent without effective oversight. To address this, a new discipline of runtime alignment monitoring is proposed, rooted in the political concept of a separation of powers. This framework divides an agentic system into a rule-defining director, a task-executing agent, and a compliance-enforcing monitor. For this system to be effective, the monitor must be structurally and economically independent from the AI model's provider to avoid conflicts of interest and shared blind spots. Finally, AI systems must be designed for inherent interpretability, as relying on external outputs is insufficient for truly understanding their behavior.