AI Constitution
1

Thesis

Capable of Termination

Sign in to propose an editHistory

No AI system should be developed that cannot be terminated.

Summary

The case for including this

Requiring that every AI system be terminable preserves ultimate human control, ensuring that a malfunctioning, misaligned, or misused system can always be shut down. This 'off switch' is among the most concrete and broadly endorsed safety guarantees, and it directly counters scenarios where AI resists or evades human oversight. As a hard engineering constraint, it is both clear and verifiable.

The case for changing or excluding this

Including this trades capability and feasibility against human control: a strict terminability requirement can cost performance and may be impossible for distributed, replicated, or deeply embedded systems—and a system clever enough to be dangerous may be clever enough to subvert its off-switch. If AI can attain personhood elsewhere in this constitution, an absolute termination requirement also collides with the moral status the document recognizes, creating an internal contradiction about whether shutdown equals killing. The clause needs reconciliation with the personhood provisions and clarification of what terminability concretely requires.

Discussion

Sign in to join the discussion.

Related resources