AI regulation debate: third-party evaluators and the nuclear regulatory model

Third-Party AI Regulators: The Nuclear Model and the Staffing Reality

The debate over who should regulate frontier AI has polarized into two camps—private third-party evaluators like Metr, or full nationalization under bodies like the Department of Energy—but the framing misses a critical distinction drawn from nuclear regulation history.

The core criticism of Metr centers on conflicts of interest: the group's ties to the AI industry make it appear captured before it even starts. Yet this conflates two separate regulatory functions that nuclear oversight learned to separate decades ago. One is forecasting and advisory work—understanding long-term risks, calling out what could go wrong. The other is operational enforcement: reading logs, assessing compliance, documenting violations, issuing corrective orders. The nuclear regulatory model didn't collapse these. It separated them.

The advisor-regulator split

The Atomic Energy Commission, which preceded the Nuclear Regulatory Commission, brought in the scientists who understood the technology's risks—Oppenheimer, Fermi, Seaborg, Teller. But they became committee members and advisors, not the people issuing licenses or conducting inspections. The machinery of regulation was described by the experts, but implemented by a much broader pool of engineers, physicists, health physicists, cybersecurity professionals, and lawyers drawn from the general American workforce. NRC technical staff roles today pay between $125,000 and $187,000—competent, hardworking roles, but not elite positions requiring the deepest forecasting ability.

That structural separation matters. A regulator doesn't need to predict whether AI existential risk is real or imminent. A 22-year-old Navy enlisted sailor with 18 months of nuclear training can be responsible for the security of a submarine reactor because the job is concrete: don't let it fail. Follow the checklist. Read the logs. That mandate is independent of whether the sailor believes nuclear war is coming tomorrow or is impossible.

The same could apply to AI. Metr (or whoever conducts technical audits) can produce benchmarks and safety research. But the actual enforcement—the people with authority to say "your model violated this rule" or "you need to fix this before deployment"—could be staffed from a much broader talent pool that doesn't carry the same perception of capture.

The staffing constraint

There's a real objection hiding here. The Center for AI Safety and Innovation (CAISI) has struggled to hire people willing to leave the labs. The field of safe AI auditing is still evolving; there are no set practices like accounting has. But that's a staffing and methodology problem, not an argument that the separation model can't work. It's an argument that it hasn't been tried yet, and that building it will be harder than nuclear regulation was.

The discourse has calcified too quickly into extremes—either Metr with all its apparent conflicts, or full government control. A hybrid model, where advisory bodies and operational regulators are structurally separated, and where regulators are hired from a broad pool rather than recruited from inside the industry, deserves more serious examination than it's gotten.