WHAT YOU NEED TO KNOW
I spent my career inside America’s nuclear enterprise, not only the command and control system, but the strategy and policy processes that kept the peace. I authored theater nuclear war plans, helped design tools for presidential nuclear decision-making, and served in the National Military Command Center in the Pentagon at the pinnacle of nuclear command and control, policy, strategy development, and war planning. I later ran for the U.S. Senate to champion peace through strength and support our current policies to re-establish it.
Command and control is only one chapter of that history. Deterrence, arms control, and non-proliferation are the others. If the dangers of uncontrolled nuclear weapons and uncontrolled artificial intelligence are comparable in the worst case, then policymakers should steal from the whole playbook, not just the briefcase.
Command and control: who is allowed to act.
Nuclear policy concentrated authority, authenticated orders, and refused to let a machine, a single technician, or an unexplained process become the last word. That lesson travels. For the most dangerous AI systems, models that can design biological threats, run autonomous cyber campaigns, or trigger strategic effects, a named human must remain accountable. Decision support is a tool. Decision replacement is a surrender of sovereignty. Require audit trails, dual control where the harm is irreversible, rigorous testing before deployment, and the legal ability to shut a system down. “The model recommended it” is not a chain of command.
Deterrence: make reckless racing unprofitable.
Nuclear peace did not come from wishing weapons away. It came from making the cost of a first strike, a surprise breakout, or a reckless test higher than the reward. AI competition with China is already an arms race in compute, talent, and deployment. Peace through strength means the United States must lead, and it means we should raise the price of unsafe breakout. That includes defending our own models and infrastructure, making sabotage of another nation’s reckless superintelligence bid a credible risk, and refusing to become dependent on closed corporate systems that can be switched off or politically constrained by executives who do not answer to the American people. Deterrence also requires credibility. Adversaries must know a human government still decides. Opacity and automation bias invite miscalculation.
Arms Control: limit the most dangerous uses, not the existence of the technology.
The useful nuclear treaties did not abolish physics. They capped deployed forces, banned certain classes of systems, created verification, and bought time. The failed early dream, the Baruch Plan’s total international ownership before an advantage existed, died because the first mover would not surrender superiority on a promise. That is a warning. Do not demand that America freeze frontier development while China scales.
Do negotiate red lines: no autonomous authorization of strategic attack; no AI-directed strikes on nuclear command, control, and warning; shared crisis communications; test-and-evaluation standards; and confidence-building measures that reduce the chance of a flash escalation. Arms control that cannot be verified is theater. Weights can be copied. Treaties must therefore focus on observable behavior, compute thresholds, military deployment, and prohibited missions, not a fantasy inspection regime for every line of code.
Non-proliferation: control the scarce inputs and the worst end uses.
The Non-Proliferation Treaty worked because weapons-grade material is hard to make and easier to watch than software. The International Atomic Energy Agency is an export-and-safeguards system, not a magic wand. AI has no plutonium. Open weights spread. That does not make non-proliferation useless. It means we should proliferate the civilian good and choke the military bad: export controls on the most advanced chips and training clusters, licensing for models above defined capability thresholds, denial of the most dangerous toolchains to hostile states and non-state actors, and an Atoms-for-Peace bargain, access to secure American compute and models in exchange for safeguards, evaluation, and refusal to weaponize. China is already flooding the world with cheap models. If we treat openness itself as the enemy, Beijing will set the global stack. If we treat every model as a warhead, we will lose the civilian race and still fail to stop the military one.
What not to import.
Nuclear weapons were state-built and state-owned. Frontier AI is not. A government monopoly on all advanced models is neither possible nor consistent with a free economy. A cartel of unaccountable labs is not a substitute for public authority either. Keep the distinction nuclear policy always made between the weapon and the underlying science: govern use, control the highest-risk capabilities, and do not pretend a general-purpose technology can be locked in a silo.
America’s constitutional system was built on individual liberty, limited government, and speech. Models trained to treat those principles as dangerous and rival ideologies as normal will export that distortion into every agency that deploys them. Safety policy that ignores that fact is not safety.
The starting point is therefore the full nuclear toolkit: human control of irreversible acts; deterrence that punishes reckless acceleration; arms control that bounds the worst military uses and creates off-ramps in a crisis; and non-proliferation that watches the chokepoints we can actually watch, compute, frontier training, and prohibited applications.
We did not keep the nuclear peace by confusing a bomber with a power plant, or a launch order with a research paper. We kept it by matching the control system to the risk. Do the same with AI. Win the race. Bind the most dangerous uses. Leave a human being on the hook.
