The AI Safety Menu

All laws

SB 1047

California SB 1047 (as vetoed). The bill everyone fought about. Pre-training protocol, kill switch, annual audits, 72-hour incidents, compute KYC.

Grows lead by~2 daysof America's lead over China, over 3 years
Cuts p(doom) by~0.33%from a 5% starting estimate
Holds up○ Training○ Lab's own use◐ Public release
Enact it on the menuSee the findings

What it does

California's 2024 bill that Governor Newsom vetoed. It required a safety plan before training, a kill switch, yearly audits, 72-hour incident reports and ID checks on cloud customers, and banned releasing a model with an unreasonable risk of catastrophe.

Applies to: Models above 10^26 FLOP costing more than $100M to train.

What's in it: Safety plan, Kill switch, Incident reports, Yearly audit, Cloud ID checks, Whistleblower protection

Where things stand

Every obligation except the audits, the KYC rules and the release prohibition was already voluntary practice at the covered labs by 2024. SB 53, enacted a year later, kept the framework, transparency and incident provisions and dropped the shutdown mandate, pre-training protocol, audits and compute-based penalties.

Why it gains ~2 days

Its release rules and audits cost about a day of U.S. progress. Its ID checks on cloud customers slow China slightly more, so on net it slightly widens the lead.

Biggest unknown: How a lab would have interpreted the duty not to release a model posing an unreasonable risk of critical harm, and whether a self-certified pre-training protocol ever delays a run.

Why it lowers p(doom) by ~0.33%

The vetoed California bill: safety plans, a kill switch, audits and incident reports.

The strongest case that it costs more

The compute-based penalty, up to 30% of training cost, creates liability that scales with ambition. A rational lab facing that exposure would evaluate more, mitigate more and release later, and the marginal caution is not captured by pricing the procedures. The undefined "unreasonable risk" standard is where the cost hides, and a court could read it in ways the lab cannot plan around. Newsom's veto message did not say the bill was expensive; it said it targeted the wrong models. The lead cost was never the strongest argument against it.

The debate

For

Against

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.