Experts: AI-Nuclear Weapon Integration Inevitable, Unclear Impact
Nuclear experts are increasingly certain that artificial intelligence will soon be integrated into nuclear weapons systems, though the precise implications of this development remain largely undefined. This consensus emerged from a mid-July gathering at the University of Chicago, where Nobel laureates convened with leading scientists, former government officials, and retired military personnel specializing in nuclear warfare. The aim of the two-day closed sessions was to inform these influential figures about the world’s most devastating weapons, ultimately guiding them to formulate policy recommendations for global leaders on averting nuclear conflict.
AI was a central topic of discussion. Scott Sagan, a Stanford professor renowned for his nuclear disarmament research, noted during a press conference that “We’re entering a new world of artificial intelligence and emerging technologies influencing our daily life, but also influencing the nuclear world we live in.” This statement reflects a widespread belief among experts that the convergence of AI and nuclear weapons is inevitable. Bob Latiff, a retired U.S. Air Force major general and a member of the Bulletin of the Atomic Scientists’ Science and Security Board, likened AI to electricity, stating, “It’s going to find its way into everything.” Latiff is also involved in setting the annual Doomsday Clock.
However, discussions about AI and nuclear weapons are complicated by fundamental ambiguities. Jon Wolfsthal, a nonproliferation expert and director of global risk at the Federation of American Scientists, points out, “nobody really knows what AI is.” Herb Lin, a Stanford professor and Doomsday Clock alumnus, echoed this sentiment, questioning, “What does it mean to give AI control of a nuclear weapon? What does it mean to give a [computer chip] control of a nuclear weapon?” He added that the debate has been largely shaped by the prominence of large language models (LLMs).
Despite these uncertainties, there is a consensus among nuclear experts on one crucial point: no one believes that consumer-facing LLMs like ChatGPT or Grok will gain access to nuclear launch codes in the near future. Wolfsthal noted that while experts hold diverse “theological” views on nuclear matters, they are united in their demand for “effective human control over nuclear weapon decisionmaking.”
Nonetheless, concerns persist regarding other potential applications of LLMs within the highest echelons of power. Wolfsthal recounted hearing proposals for interactive computer systems that could help a president predict the actions of adversaries like Putin or Xi by analyzing their past statements and writings. He questioned the reliability of such systems, noting, “How do you know Putin believes what he’s said or written?… it’s just based on an assumption that can’t be tested.” He also expressed skepticism that those developing such tools fully grasp the decision-making environment of a president.
The U.S. military is already exploring AI’s role in nuclear operations. Last year, Air Force General Anthony J. Cotton, who oversees America’s nuclear forces, spoke about the importance of AI adoption, stating that nuclear forces are “developing artificial intelligence or AI-enabled, human led, decision support tools to ensure our leaders are able to respond to complex, time-sensitive scenarios.”
Wolfsthal’s primary concern is not a rogue AI initiating nuclear war, but rather the vulnerabilities created by automating parts of the nuclear command and control system. He fears such automation could be exploited by adversaries or produce data and recommendations that humans are ill-equipped to understand, leading to disastrous decisions.
Launching a nuclear weapon is a highly complex process involving intricate webs of early warning radar, satellites, and other computer systems, all monitored by humans. For instance, an American nuclear launch requires multiple human actions, including two individuals turning keys simultaneously in a silo. The entire process is the culmination of numerous human decisions.
The prospect of AI taking over parts of this process raises critical questions. What happens when an AI, not a human, is monitoring early warning radar? U.S. nuclear policy requires “dual phenomenology” to confirm a nuclear strike, meaning an attack must be verified by both satellite and radar systems. Wolfsthal argues that, at this stage, AI cannot serve as one of these confirming phenomena. This is partly due to the “black box” nature of many AI systems, whose internal workings are not fully understood. Experts generally agree that integrating such systems into nuclear decision-making would be ill-advised.
Latiff also voiced concerns about AI systems reinforcing confirmation bias. He worries about the true meaning of “human control” if AI heavily influences decisions. “I’ve been a commander,” he stated. “I know what it means to be accountable for my decisions. And you need that… If Johnny gets killed, who do I blame?” AI systems, being bound by their programming and training data, cannot be held responsible for failures and are inherently limited in their ability to “see outside themselves.”
Herb Lin brought up the critical example of Stanislav Petrov, a Soviet Air Defence Forces lieutenant colonel who, in 1983, disregarded a machine alert indicating a U.S. missile launch, thereby averting potential nuclear war. Petrov’s judgment, based on his experience and intuition (e.g., the small number of missiles indicated, the newness of the computer system), overrode the machine’s apparent error. Lin questions whether humans can be routinely expected to make such critical judgment calls that go “outside your training data,” an ability AI, by definition, lacks.
Meanwhile, the U.S. government has prioritized AI development, often framing it in terms of a new arms race. In May, the Department of Energy proclaimed on X that “AI is the next Manhattan Project, and the UNITED STATES WILL WIN,” explicitly likening the pursuit of AI to a competition with China. Lin found such metaphors “awful,” noting that the Manhattan Project had a clear, tangible success metric (the explosion of a nuclear weapon), whereas the success of an “AI Manhattan Project” remains undefined.