28 of 43 AI safety laws cost America less than a week of its lead over China

We priced 43 proposed AI safety laws by what each would cost America's lead over China over 3 years, and how much each lowers p(doom), the chance of an AI catastrophe. Assuming China doesn't depend on U.S. models, the case least favorable to safety laws.

Gains lead2 laws

Cloud ID checks, Lock up the weights

Less than a day16 laws

Rulebook with a floor, Background checks, DNA order screening, Humans control nukes, Incident reports, Transparency report, Yearly audit, Outside safety tests, Safety tests, Safety plan, 4-week test period, U.S.–China safety deal, Limits on AI surveillance, Whistleblower protection, Red team if risky, Fund defensive tech

A day to a week10 laws

Kill switch, Keep weights closed, Power-grab safeguards, Prove it's safe before launch, Chips that enforce rules, Checks before lab use, Clearances for lab staff, Liability insurance, Alignment moonshot, International AI agency

A week to a month7 laws

Liability for AI harms, 30-day training notice, No capabilities papers, Air-gapped labs, Pause with China, Limit AI-run research, Limit AI agents

A month or more8 laws

U.S. 6-month pause, License to train, A third of R&D on safety, Global superintelligence ban, Global pause, Global compute cap, U.S. compute cap, U.S. superintelligence ban

5% → ~2.8%p(doom) with just the 28 laws that cost less than a week.

All 8laws costing a month or more hold up training itself: pauses, compute caps, licenses and bans. Most other laws hold up only public release, or nothing, so labs keep building.

~0.88%p(doom) with all 43 safety laws, mostly thanks to deals with China, if they hold.

Play the menu See every law

Rough starting points, not precise forecasts. Every number, source and assumption is on the menu, where you can change them. p(doom) starts from the 5% median of AI researchers surveyed.