AI Can Play a Limited but Crucial Role With Nuclear Weapons

There is much discussion about the dangers of artificial intelligence and nuclear weapons these days. Indeed, the Vatican recently hosted a global conference on the topic that featured Nobel-prize winning scholars, experts on AI, and proponents of nuclear disarmament debating the threat.

will introduce destabilizing dynamics among nuclear-armed states, particularly by diminishing direct human involvement in decisions about nuclear weapons use. Yet many of the same experts also recognize potential stabilizing effects in certain contexts.

On balance, however, the risks to strategic stability from significantly accelerating nuclear decision timelines or reducing human involvement in launch decisions and execution are likely to outweigh the potential benefits.

That same year Gen. Anthony Cotton, the commander of U.S. Strategic Command, testified to the Senate that “USSTRATCOM will use AI/ML to enable and accelerate decision-making.… [but] AI will remain subordinate to the authority and accountability vested in humans.”  

It is clear that responsible nuclear actors cannot cede to a computer the decision to employ nuclear—but the Department of War SHOULD integrate AI and large language models to help it understand escalation pathways and how adversaries may react to various courses of action during times of acute conflict.

For example, should the United States find itself in a high-intensity conflict with a nuclear-armed adversary such as China, Department of War planners must think carefully about how far it can prosecute the war without inadvertently triggering a nuclear response from the adversary. Planners may find striking strategic targets with conventional weapons appealing, particularly if it can help them achieve their war aims—but would be rightfully concerned about pushing an adversary too far and triggering a nuclear retaliation.

Similarly, the United States may feel compelled to employ nuclear weapons to achieve operational objectives, respond to a Chinese nuclear use, or end the conflict decisively on terms favorable to the United States—but do so in a way that does not risk a general nuclear exchange between the United States and China.

An AI model could be useful in any of the above scenarios in helping war planners understand the likelihood of potential Chinese reactions to American strike options. If the Department of War had AI programs that modeled Chinese decision-making and individual Chinese leader profiles, assumed Chinese strategic objectives and current order of battle among all combatants, and a maintained sense of perceived “stake” for the combatants, such a model could come up with a likelihood of various Chinese responses. 

For instance, such a model could come up with results suggesting that, in response to the United States destroying a naval base on Hainan island using conventional munitions, there would be a 68% chance that China would respond with large-scale conventional attacks on American bases in Guam and Japan, a 20% chance they would do nothing, a 10% chance they would destroy the American base in Guam with a nuclear weapon, and a 2% chance they would strike Hawaii with a nuclear weapon.

Alternatively, a military commander may be ordered by the President to carry out a nuclear strike on Chinese forces in response to a Chinese nuclear first strike—but to do so in a way that would not further escalate the conflict and ideally deescalate the conflict. The commander could then order a series of strike options be considered by the AI model, which would then offer a series of calculations on which option would most likely terminate the conflict and which option stood the highest chance of precipitating a full-scale nuclear response from the adversary. Military commanders would then use this model as a single input when deciding courses of action.

This is not to say that AI should have authority to determine U.S. military courses of action or target sets—humans should never cede such ground. But it can be one input to help military leaders and elected officials think through potential enemy reactions during moments of the most acute crisis.

In short, nuclear strategy fundamentally is about anticipating how an adversary will respond in extreme and largely unprecedented contingencies. If AI models can predict that better than humans can, then it's irresponsible not to use them.

Robert Peters is the acting Director of the Heritage Foundation’s Allison Center for National Security.

image: A B-52H Stratofortress is parked on the flight line during Exercise Global Thunder 26 (GT26) at Minot Air Force Base, North Dakota, Oct. 23, 2025. Strategic deterrence includes a triad of nuclear delivery platforms and associated support systems; an intelligence and sensing apparatus to give indications and warnings of incoming threats; assured national and nuclear command, control and communications; the necessary infrastructure to sustain and maintain reliable warheads; a credible missile defense. (U.S. Air Force photo by Airman 1st Class Wesley Davies)