The Intersection of AI, Government, and Uncertainty

There is a fundamental question emerging around artificial intelligence that we have not yet adequately answered: What is the proper role of government when nobody knows where a technology is going?

If there were ever a time for government to get involved early, this is it. Not because we know AI will become dangerous, and not because government can predict or control the future. It is because we don’t know what the risks are—and the consequences of being wrong could be extraordinarily high.

Government intervention is usually justified by evidence that a problem already exists. But when the probability of a catastrophic outcome is deeply uncertain, the question is not simply “How likely is it?” It is: “What happens if we’re wrong?”

The debate often divides into polarized camps. Technologists point to immense upsides: better medicine, scientific breakthroughs, and historic productivity gains. Others warn of darker scenarios: autonomous weapons, destabilized societies, mass unemployment, and humans losing control of autonomous systems. Still others argue that focusing on remote, catastrophic scenarios distracts us from immediate harms.

Suppose we genuinely do not know who is right. If the downside of an error were trivial, uncertainty would be a good reason to let the market experiment. But even a minute chance of systemic catastrophe fundamentally changes the calculation. The goal should not be to stop AI or control every aspect of its development. It should be to reduce the probability and consequences of the worst outcomes while preserving as much of the enormous upside as possible.

In confronting that challenge, we have an important precedent: the international effort to control nuclear proliferation.

Nuclear weapons showed the world that when a technology can produce catastrophic consequences, governments cannot simply wait for certainty before acting. In 1945, scientists and statesmen could not foresee what an atomic world would look like. But after the Cuban Missile Crisis brought humanity frighteningly close to war in 1962, leaders recognized that waiting for catastrophe was not an acceptable strategy.

The United States and the Soviet Union could not eliminate the technology, nor could they eliminate their rivalry. Instead, they focused on risk reduction. They established direct communications channels, negotiated verification protocols, and created the 1968 Nuclear Non-Proliferation Treaty. It was imperfect, but perfection was never the standard. Risk reduction was.

Crucially, the lesson was never to regulate everything. Nuclear physics was not banned; nuclear energy and scientific research moved forward. What governments controlled were the narrow conditions creating extraordinary risk: weapons proliferation, unauthorized use, and accidental escalation.

AI requires that exact distinction. There is no reason for government to dictate ordinary applications or stop people from using AI to write, code, or design products. The real question is not “Can we regulate AI?” It is: “Which capabilities and applications create risks so consequential that society cannot reasonably leave them entirely to private actors?” That is a far more manageable challenge.

Containing AI, however, is unique. Nuclear weapons require rare physical materials and vast facilities that leave clear footprints. Software, by contrast, crosses borders instantly; models can be copied, and knowledge is distributed. We cannot put AI in a vault, nor draw a line on a map and expect it to stop at the border.

Yet frontier AI still depends on an inescapable physical footprint. Cutting-edge foundation models rely on physical choke points: specialized semiconductor manufacturing, global supply chains, and massive data centers demanding enormous electrical power. These physical dependencies offer realistic touchpoints for oversight, safety standards, and tracking without burdening ordinary development.

The most difficult obstacle is the security dilemma. Country A accelerates development out of fear of falling behind Country B, and Country B accelerates faster in response. Neither wants to build dangerous systems, but neither wants to be the one that slows down. In a classic collective-action trap, unilateral restraint feels impossible.

This is why domestic guardrails must be paired with international diplomacy. Domestically, governments must build deep technical competence. We should never want politicians dictating algorithms, but governments must be capable of evaluating frontier capabilities, mandating rigorous pre-deployment testing, clarifying legal liability, and safeguarding critical infrastructure.

Internationally, the goal is not total trust. Cold War adversaries certainly did not trust each other. The goal is enough transparency, shared rules, and verifiable redlines—beginning with non-negotiable human control over nuclear command systems—so that neither side is surprised into catastrophe.

The most dangerous assumption is that we have plenty of time. Waiting until an AI system causes an irreversible disaster before establishing safeguards means acting too late. Once unvetted, autonomous systems become deeply embedded in defense grids, financial markets, and vital infrastructure, changing course will be nearly impossible.

The argument for government begins with humility. Nobody knows what AI will become. That is precisely the reason to be involved. Government should not pretend to pick economic winners or micromanage innovation. But democratic governance has a basic responsibility to ask what risks we are willing to accept when the cost of being wrong could be beyond measure.

The real question isn’t whether government can control AI—it probably can’t. The question is whether governments can act early enough to reduce the consequences of being wrong. That may be the most important AI policy question of our time.