The Atom Did Not Choose

In July 1945, the Trinity test in New Mexico produced a light brighter than anything humans had previously made, followed by a shockwave that knocked observers down miles away. J. Robert Oppenheimer reportedly thought of the Bhagavad Gita: "Now I am become Death, the destroyer of worlds." Within weeks, two Japanese cities were erased. Within years, nuclear power plants were producing electricity for civilian use. The same physical process. Radically different applications. The determining factor was not the technology — it was every choice made by every government, scientist, engineer, and military commander about what to do with it. This is the pattern that matters when thinking about artificial intelligence. Not as a warning that AI will destroy cities, but as a clear example of how a technology's character is not fixed. It is determined by the choices of the people and institutions that develop and deploy it.

Dual Use Is Not an Accident

The nuclear analogy gets dismissed sometimes because nuclear weapons were deliberately developed as weapons before peaceful applications were pursued. AI development runs in the opposite direction — most foundational research is aimed at beneficial applications, and the dangerous uses are secondary or unintended. That inversion actually makes the analogy more relevant, not less. When dangerous capabilities emerge as byproducts of beneficial ones, the governance challenge is harder. You cannot simply classify the dangerous applications and contain them, because they are inseparable from the capabilities that produce the benefits. A language model trained to generate helpful, accurate text is also capable of generating persuasive false text. A system trained to identify patterns in biological data to accelerate drug discovery has applications in understanding pathogens that could be engineered for harm. The capability does not separate cleanly along lines of intent.

The Governance Gap Is the Real Problem

Nuclear technology generated a governance response — imperfect, contested, and never fully adequate — but real. The Nuclear Non-Proliferation Treaty, the International Atomic Energy Agency, export controls, and decades of arms reduction negotiations represent a sustained, if incomplete, attempt to manage a dual-use technology through international coordination. Research from the Stockholm International Peace Research Institute examining technology governance regimes has found that the speed of capability diffusion is the primary variable predicting governance failure. Technologies that spread to many actors before governance structures are in place systematically produce worse outcomes than technologies where governance development and capability development proceed together. AI capabilities are diffusing faster than nuclear capabilities ever did, to more actors, with lower technical barriers. The governance gap is correspondingly larger.

The Tangent: What We Got Right With Nuclear

For all the failures of nuclear governance — arms races, near-misses, continued proliferation risks — there are things that went better than they might have. Nuclear weapons have not been used in conflict since 1945, a period covering dozens of major wars. That is not a trivial achievement. It resulted from a combination of deterrence theory, arms control agreements, non-proliferation norms, and institutional investment in stability. The people who built those structures made consequential choices about what kind of nuclear world they wanted to live in. They were not optimistic about human nature or the reliability of any particular government. They were pessimistic in productive ways — they designed systems that assumed mistakes would happen and built in mechanisms for managing consequences when they did. That kind of structured pessimism may be exactly what AI governance needs. Not an assumption that bad actors will behave well, but systems designed to manage the world where they do not.

Who Is Doing the Equivalent Work Now

Some institutions are attempting the equivalent work for AI. The pattern of that work is fragmented, underfunded relative to AI development itself, and largely voluntary in its commitments. Research from Georgetown University's Center for Security and Emerging Technology on AI governance institutions has documented the gap between the resources going into AI capability development and the resources going into AI safety and governance research — a ratio that runs in the hundreds to one in capability's favor. This is not a sustainable ratio if we take the analogy seriously. Nuclear weapons programs were accompanied by significant investment in understanding the consequences of nuclear use and how to prevent escalation. The equivalent for AI has not materialized at scale.

The Pattern Requires Active Choice

Nuclear technology did not automatically become a tool for powering cities rather than only destroying them. That required active choices by governments, engineers, and institutions to invest in peaceful applications and build regulatory frameworks for safe use. The same is true for AI. The dual-use pattern does not resolve itself. It requires people who see the full picture — the power and the danger, the benefits and the risks — and who make decisions accordingly. Those people exist in AI right now. Whether they have enough institutional support to do the necessary work is a different and more uncertain question.