Frontier Model Regulation

The FAA analogy

Dario Amodei argues frontier AI models should be regulated the way airplanes, drugs, or automobiles are: powerful technologies a modern economy genuinely depends on, but ones capable of mass harm if poorly designed or operated. The FAA model, pre-release testing paired with the power to block an unsafe aircraft, is the template he proposes, rather than the FDA's after-the-fact approval process or a banking-style regime of continuous supervision.¹

The proposal scopes testing to four specific, testable risk domains rather than a broad and vague mandate. Cybersecurity: can the model materially enable attacks on critical infrastructure, financial systems, or national-security targets. Biological weapons: can the model provide meaningful uplift to someone trying to create one. Loss of control: does the model show autonomous behavior that evades human oversight. Automated research and development: can the model accelerate AI capability development in ways that compound the other three risks. Testing is deliberately scoped to just these four, on the argument that a broad testing mandate creates compliance theater, where resources flow to low-risk checklist items while the genuine risks go unaddressed, alongside explicit protections against political favoritism in blocking decisions. Under the proposal, models above a defined compute threshold would undergo mandatory testing by an independent, technically capable third party, with government holding the power to block or delay deployment of anything that fails.

From transparency to binding rules

Amodei's own position moved in stages. Through 2023 and 2024, Anthropic advocated transparency legislation, disclosing safety procedures, test results, and critical safety incidents, on the reasoning that the specific shape of AI risk was still unclear and that rigid legislation written too early would likely be wrong in some costly way; transparency was meant to build the evidence base for smarter regulation later. The public emergence of cyber capability at nation-state-relevant scale in a frontier model changed that calculus in his account: the shape of the risk stopped being theoretical, which is his stated reason for moving from transparency to binding regulation. He is explicit that Anthropic's own voluntary internal scaling policy is not a substitute for the legislative floor he is proposing, since a voluntary framework can adapt to new evidence, but it can also simply be changed or abandoned, which is exactly what a mandatory floor is meant to prevent.

A second, blunter lever

Pre-release testing governs whether a frontier model may be deployed at all. A separate legal tool, export control, governs who may access a model that has already cleared that bar, and it operates very differently: it can be applied to cut off specific groups, including a lab's own foreign-national employees, from its newest models, and determinations made under it are exempt from judicial review, a faster and less checkable instrument than the FAA-style testing regime, aimed at controlling diffusion and access rather than qualifying safety.