AI shutdown claims: six questions for separating policy, control and evidence

[noqt]

‘AI kill switch’ can describe very different things: a policy direction, a proposed emergency authority, a bounded risk assessment, or an incident intervention. Treating those as one control claim makes governance discussions less precise.

We compared four current public sources—California Executive Order N-9-26, introduced H.R. 9917, METR’s May 2026 frontier-risk assessment, and OpenAI’s first-party Hugging Face incident report—using six questions:

• Authority: who can issue and authenticate the order?
• Scope: which systems, workloads or users does it reach?
• Actuation: what action can actually be taken?
• Independence: who verifies the result beyond the operator?
• Evidence: what observable proof survives the action?
• Recovery: what is preserved, restored or rebuilt afterwards?

For governance work, the distinctions matter: proposed is not failed; observed is not independently verified; and a source not publicly establishing a control is not proof the control does not exist. The comparison also keeps throttling, access suspension, inference stops, shutdown and fallback from being treated as interchangeable.

The source-linked claim card is here:
https://noqt.no-qt.chatgpt.site/research/ai-shutdown-claim-card

This is a noqt Research source comparison, not legal advice, a compliance framework, a product test or evidence of a universal AI shutdown capability. If a row overstates a source, the useful correction is the claim and the relevant primary passage.