The AI Safety Menu

All laws

Incident reports

Critical safety incident reporting. Tell the state within 15 days when something goes seriously wrong. 24 hours if it is imminent.

Lead costAlmost noneof America's lead over China, over 3 years
Cuts p(doom) by~0.05%from a 5% starting estimate
Holds up○ Training○ Lab's own use○ Public release
Enact it on the menuSee the findings

What it does

When something goes badly wrong, such as stolen model weights or a model helping to cause serious harm, the lab reports it to the state within 15 days. If lives are at immediate risk, it has 24 hours.

Applies to: Frontier developers (models above 10^26 FLOP).

Where things stand

OpenAI and Anthropic already have incident processes under their own frameworks, and both are now legally bound by SB 53 in California. Nothing in the requirement touches training, internal use or release.

Why it costs almost no lead

A report is filed after an incident, so it never holds up a model. The staff time displaces almost no research, so the effect on the lead is under a day.

Biggest unknown: None worth naming. The only cost is a lawyer's afternoon.

Why it lowers p(doom) by ~0.05%

Lets the government spot dangerous failures early, like a stolen model or AI helping with an attack.

Reports come after something has gone wrong, so they mostly help regulators respond and learn. They rarely prevent the first disaster.

The strongest case that it costs more

A reporting regime creates a record that regulators and plaintiffs can use later, which could make labs more cautious about internal deployments of capable models. That caution, not the filing, is where a cost would come from. We see no evidence of it so far and it would be a cost of liability exposure rather than of the reporting rule itself.

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.