ITHome snapshot and original · Published 2026-09-17
What can an AI emergency stop do? Several perspectives
The disagreement over an emergency stop is really about placement: is it one layer among many, or a stand-alone safety device?
Four people, four reservations
Hinton's doubt is long-term: a superintelligence could plausibly talk its overseer out of pressing the switch. Soares's doubt is about what lies beyond the boundary: once AI can copy itself or reach critical infrastructure, there is no guarantee the switch still binds.
Jack Clark and Dario Amodei represent the other side: Clark argues for mandatory, independently verifiable shutdown mechanisms; Amodei calls a switch far from a panacea, one part of a broader strategy of testing, external evaluation and deployment limits.
From a device to a system
Engineering today can deactivate self-hosted models and cut tool or compute access; the contested ground is systems beyond that perimeter. The policy question accordingly shifts to mandated retention plus independent verification rather than a single button.
The backdrop is Sam Altman's 2024 remark that there is no magic red button — safety comes from a series of concrete decisions. The debate's value is turning an abstract risk into a checkable list of mechanisms.
Sources and notes
sources中文
Sources and notes
- Read the originalITHome snapshot and original
Retained summary and context
The positions are quoted via Business Insider's report, observed here through ITHome's Chinese coverage; a stated position is not a shipped mechanism, and the debate itself is not proof of the risk's size.
Sources and evidence · 2 links
Each link has a purpose and scope. Link counts do not establish independent confirmation; direct support applies only to the statements specified below.
Source timeline
- · Entered the BitShovel radar: 60 entries in the current ITHome feed.
Collection record
First discovered:
Collection and source publication have distinct dates.