The AI Safety Menu

All laws

Global superintelligence ban

Verified global ban on superintelligence. A treaty, with inspections, banning superintelligence everywhere until it can be built safely.

Cuts lead by~3.5 monthsof America's lead over China, over 3 years
Cuts p(doom) by~1.5%from a 5% starting estimate
Holds up● Training◐ Lab's own use○ Public release
Enact it on the menuSee the findings

What it does

The U.S., China and other countries would agree not to build AI that is smarter than humans at almost everything until scientists agree it can be done safely. Inspectors would track AI chips and large data centers so that no country could secretly train such a system.

Applies to: Frontier AI development in every signatory country.

Where things stand

There is no treaty, only petitions, open letters and draft texts written by nonprofits. The UN's new AI scientific panel and Global Dialogue are advisory and cannot restrict development.

Why it costs ~3.5 months

Both countries stop building the most powerful systems, so most of the cost cancels. We assume verification catches most cheating but not all, so China keeps about 30% of its progress. Net: about a month a year in.

Biggest unknown: Whether the U.S. and China could verify each other's compliance well enough that neither suspects a secret program.

Why it lowers p(doom) by ~1.5%

Stops every country, China included, from building AI smarter than humans until it can be done safely.

If it holds, it removes the race to build the systems people worry about most. Whether it holds depends on verification that doesn't exist yet, so our central estimate is well below its full potential and the range is enormous.

The strongest case that it costs more

Verification tools for AI are unproven. A treaty that China quietly breaks leaves America both behind and no safer, which is the worst corner of the map.

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.