The AI Safety Menu

All laws

International AI agency

International AI agency with inspections. An IAEA for AI: an international body that inspects frontier labs and checks safety standards.

Cuts lead by~6 daysof America's lead over China, over 3 years
Cuts p(doom) by~2 pointsfrom a 5% starting estimate
Holds up◐ Training○ Lab's own use◐ Public release
Enact it on the menuSee the findings

What it does

A new international body would be allowed to send inspectors into frontier AI labs and data centers, audit the most powerful models and check that safety rules are followed, much like IAEA inspectors visit nuclear plants. Labs above a set size in member countries would have to open their doors.

Applies to: Frontier AI developers in member countries.

Where things stand

Government testing is national and mostly voluntary (U.S. CAISI, UK AI Security Institute). UN bodies only assess and convene; the 2024 UN advisory body report chose an IPCC-style panel instead of an agency.

Why it costs ~6 days

Inspections take staff time in both countries and occasionally hold up a launch. Because China is inspected too, most of the cost cancels. Net: about 2 days.

Biggest unknown: Whether the U.S. and China would ever let foreign inspectors into their top labs, including military and classified AI work.

Why it lowers p(doom) by ~2 points

Puts inspectors inside frontier labs in every member country, the way the IAEA inspects nuclear sites.

An inspector in the room makes hidden dangerous projects much harder. It works only if major powers let inspectors in and act on what they find.

Set to match expert forecasters' estimate for approval of frontier models by an international body, the closest policy they rated (Forecasting Research Institute, Oct 2026).

The strongest case that it costs more

Inspectors could see U.S. trade secrets, and an agency with real power would take years to set up. Weak inspections could give false comfort.

What expert forecasters think

The closest policy they assessed, approval of frontier models before release by an international body including the U.S. and China, cut forecasts of an AI catastrophe by 2050 by 23% to 50% (experts: 2% to 1%). Experts give it a 5% chance by 2030.

Forecasting Research Institute, Oct 2026

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.