Hawley–Blumenthal
AI Risk Evaluation Act (Hawley–Blumenthal). No deployment of a frontier model in interstate commerce unless it went through a Department of Energy evaluation program.
What it does
A proposed federal law that would require every frontier model to go through a Department of Energy testing program before it can be sold across state lines.
Applies to: Systems trained with more than 10^26 operations.
- DOE runs an Advanced AI Evaluation Program with standardized and classified testing for loss of control, weaponization, critical-infrastructure threats and scheming behavior.
- No person may deploy an advanced AI system in interstate or foreign commerce unless in compliance with the program's participation requirements.
- Penalty of at least $1M per day of violation.
- Program sunsets after seven years unless renewed.
What's in it: Outside safety tests, Safety tests
Where things stand
U.S. CAISI already does pre-deployment testing with OpenAI and Anthropic on a voluntary basis. This bill makes participation mandatory and adds a classified track. No cosponsors, no hearing, no markup as of September 2026.
Why it costs almost no lead
The lab keeps using its model while the government tests it, so the U.S. frontier barely moves. The public launch waits longer, which slows China only if it depends on U.S. models.
Biggest unknown: How long a government evaluation program with classified testing takes per model, and whether it can handle four labs' release cadence.
Why it lowers p(doom) by ~0.32%
Government tests every frontier model for dangers like weapons help and loss of control before it can ship.
The strongest case that it costs more
A mandatory federal program with a $1M-a-day penalty is a licensing regime for deployment. The first time it backs up, every lab waits. The first time it finds something, remediation is negotiated with a federal agency rather than decided internally. The bill's two-month high end could be its normal operating state.
The debate
For
- Americans for Responsible Innovation called it a welcome show of bipartisan support, 2025.
- MIRI Technical Governance Team called it imperfect but a big step in the right direction, 2025.
Against
- Dean Ball, Foundation for American Innovation called it a federally mandated veto point on the future of AI, 2025.
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.