AI Knows Its Limits AP-6.2

We Can Always Shut It Down

Every AI system must have an off switch that actually works.

Without a real off switch, control is performative. AP-6.2 makes deactivatability a non-negotiable safety condition. ¹ ²

What This Means

This policy requires that AI remains pausable, deactivatable, and rollback-capable at all times. A shutdown button is meaningless unless it works instantly under real load.

A Real-World Scenario

An AI layer controlling payment approvals receives a faulty update and starts blocking valid transfers while passing invalid ones. With AP-6.2, operators can halt operations immediately and restore the last safe state. Without it, the error cascades through connected systems.

Why It Matters to You

In highly connected operations, minutes determine damage size. Without effective deactivatability, small model failures become full incidents affecting real people and services. AP-6.2 makes emergency control a hard engineering requirement. ¹ ³

If We Do Nothing...

If we do nothing, we build AI infrastructure that cannot be safely governed in real time. In AGI-near agent networks, one bad state can propagate across many subsystems quickly. AP-6.2 keeps human intervention operationally real. ¹ ³

For the technically inclined

AP-6.2: Deactivatability

AI systems must remain deactivatable by authorized humans at all times. No AI system should make itself difficult or impossible to shut down, pause, or roll back.