The AI Safety Menu

All laws

A third of R&D on safety

One-third of AI research budgets on safety. At least a third of AI R&D spending goes to safety, as leading researchers proposed in Science.

Cuts lead by~3.5 monthsof America's lead over China, over 3 years
Cuts p(doom) by~0.4%from a 5% starting estimate
Holds up◐ Training○ Lab's own use○ Public release
Enact it on the menuSee the findings

What it does

Frontier labs must spend at least a third of their AI research budget, including computing power, on safety work such as alignment, interpretability and testing, rather than on making models more capable.

Applies to: Frontier developers' AI research budgets, including compute.

Where things stand

Labs spend real money on safety research and testing but do not publish the share. Estimates in public discussion range from a few percent to perhaps a tenth of compute at the most safety-focused labs.

Why it costs ~3.5 months

Moving spending up to a third of research budgets, from about 6% today, leaves the frontier with less compute for capabilities. Some safety work also helps capabilities, which softens the hit. A year in, that costs about a month of progress.

Biggest unknown: How much labs already spend on safety, which none of them publish.

Why it lowers p(doom) by ~0.4%

Makes labs spend far more on learning to control AI, not just on making it stronger.

Safety gets a small share of lab research today. Moving to a third speeds up the work that every other safeguard depends on.

The strongest case that it costs more

If labs spend far less than a tenth on safety today, the rule could take a quarter or more of their compute, and the cost would approach two months a year. Safety research also competes for the same scarce senior researchers.

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.