Kill switch

Full shutdown capability. Be able to turn the model off. The kill switch.

Cuts lead by~1 dayof America's lead over China, over 3 years

Cuts p(doom) by~0.1%from a 5% starting estimate

Holds up○ Training○ Lab's own use○ Public release

Enact it on the menu See the findings

What it does

The lab must be able to stop a training run quickly and switch off any model it controls, and prove it can.

Applies to: Covered frontier models (SB 1047 definition); under study for California by executive order.

Implement the capability to promptly enact a full shutdown of training, model operation and derivatives under the developer's control (SB 1047 § 22603(a)).
California's September 2026 executive order asks an expert panel to recommend how a kill switch for frontier models should work.

Where things stand

Every lab can stop a training run and pull a model from its API. The ability to do so is operational hygiene, not a new engineering program. Anthropic's RSP and OpenAI's frameworks both contemplate pausing deployment as a mitigation.

Why it costs ~1 day

Labs can already stop runs and pull models from their apps. Documenting and testing that ability costs under a day of progress.

Biggest unknown: Whether a regulator's definition of shutdown ever exceeds what labs already engineer for their own operations.

Why it lowers p(doom) by ~0.1%

Makes sure a lab can actually stop a model that goes wrong.

A kill switch helps only if someone notices the problem in time and the model hasn't already spread, so the effect is modest.

The strongest case that it costs more

A shutdown requirement that extends to model derivatives held by third parties, or to open-weight models, is not technically achievable and could push a lab to withhold open weights. That would affect China's access more than the U.S. frontier, which would push the price negative rather than positive.

The debate

For

Gov. Gavin Newsom ordered experts to advance a kill switch for frontier models, September 2026.
Reps. Ted Lieu and Nathaniel Moran introduced the AI Kill Switch Act, July 2026.
The AI Policy Network backed the Kill Switch Act and cited polling that 86% of voters support off-switches, 2026.
California SB 1047 would have required full shutdown capability, 2024.

Against

Fei-Fei Li argued SB 1047's kill switch would make developers hesitant to write and share code, 2024.
Adam Thierer, R Street argued the Kill Switch Act won't stop rogue AI but will slow innovation, 2026.

Sources

Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.