The AI Safety Menu

All laws

FRONTIER Act

FRONTIER Act (Obernolte–Trahan). A federal SB 53 with audits, a Commerce under secretary for AI security, and preemption of state developer rules.

Cuts lead by~3 daysof America's lead over China, over 3 years
Cuts p(doom) by~0.14%from a 5% starting estimate
Holds up○ Training○ Lab's own use◐ Public release
Enact it on the menuSee the findings

What it does

A proposed federal law that sets California-style rules nationwide, adds outside audits, and stops states from writing their own versions.

Applies to: Frontier models above 10^26 FLOP; large developers ($50M revenue and $1B AI spend) and very large developers ($5B and $10B).

What's in it: Safety plan, Yearly audit, Transparency report, Incident reports

Where things stand

Everything in it is already practiced voluntarily or required by SB 53 and RAISE, except the six-monthly independent verification assessments, which resemble the audits SB 1047 proposed.

Why it costs ~3 days

Mostly paperwork plus audits. The model estimates about a day of lost U.S. progress.

Biggest unknown: Whether the six-monthly independent verification assessments for very large developers ever become a release gate in practice.

Why it lowers p(doom) by ~0.14%

A federal version of California's law, with audits, that would override state rules.

The strongest case that it costs more

An under secretary with emergency-order power and a $1M-a-day penalty is a regulator that can stop a launch by letter. The statute does not say when that power is used. The price here assumes it is used rarely; a critic would say federal emergency powers, once created, are used.

The debate

For

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.