Effective military A.I. enables a credible first-strike capability, upending MAD, creating use-it-or-lose-it pressures and pushing adversaries towards launch-on-warning postures. This ensures accidental wars and nuclear escalation.

In an extreme case, AI could undermine the condition of MAD and make nuclear war winnable, but it takes much less to undermine strategic stability. AI advancements merely need to cast doubt on the credibility of retaliation at some level of conflict. Major nuclear powers, such as the United States, Russia, and China, have a shared interest in maintaining the credibility of central deterrence, but they seek regional advantages in pursuit of what they regard as their core strategic interests. Areas where credibility is already strained, such as certain extended deterrence guarantees, are particularly vulnerable to destabilization. The increasingly multipolar strategic environment is also encouraging forms of competition that threaten stability. For instance, the United States is interested in developing the capability to track and target a minor nuclear power’s mobile missile launchers, but Russia and China fear that the same technology could mature into a threat to their more sophisticated retaliatory forces. In a crisis situation, the employment or availability of AI-enabled intelligence, surveillance, and reconnaissance (ISR) or weapon systems could stoke tensions and increase the chances of inadvertent escalation. Finally, the pursuit of advanced military capabilities is liable to cause arms race instability even if those technologies are nonviable, as in the historical case of missile defense.

The challenge AI poses to strategic stability is not unique to this particular technology, but it is more acute because of rapid technical progress in AI and its many potential intersections with nuclear strategy. Most of the specific applications AI are likely to be used for, such as analysis of ISR data, controlling autonomous sensor platforms, and automated target recognition (ATR) have been eagerly sought for decades but were beyond the capability of available technology. Even without further breakthroughs, incremental progress using existing AI techniques may make these long-sought goals practical realities in the foreseeable future.

Both Russia and China appear to believe that the United States is attempting to leverage AI to threaten the survivability of their strategic nuclear forces, stoking mutual distrust that could prove catastrophic in a crisis. As Paul Bracken observes, ongoing improvements in technology such as AI threaten to “undermine minimum deterrence strategies” and “blur the line between conventional and nuclear war” (Bracken, 2017).

AI in the Cold War

AI pioneer Marvin Minsky defined AI as “the science of making machines do things that would require intelligence if done by men” (Minsky, 1968, p. v). Since AI research began in the 1950s, the boundaries of the field have shifted as computers have reshaped how humans comprehend “intelligence.” AI has also evolved as theoretical paradigms have shifted in and out of vogue. From the 1950s until the 1980s, a “symbolic” paradigm that aimed to replicate high-level human reasoning predominated, only to be supplanted by a “connectionist” paradigm that sought to emulate the biological basis of human cognition using artificial neural networks. In the 20th century, neither paradigm worked particularly well outside laboratory demonstrations. This triggered occasional periods (sometimes characterized as AI winters) during which funding for AI research was scarce. Thanks to decades of progress in computer science, advances in computing and communications hardware and software, and the rise of cloud computing and big data, AI has advanced rapidly in the past few years, most prominently in the field of “deep neural networks” (DNNs), or neural networks with many layers (Goodfellow, Bengio, and Courville, 2016). The increase in the performance of DNNs has been so spectacular that they have become almost synonymous with AI, but in actuality the older paradigms are also continuing to progress and are in widespread commercial and military use. Some impressive recent AI systems, such as Alphabet DeepMind’s AlphaGo program, which beat the world Go champion, employ DNNs in combination with such older techniques as searching the tree of possible moves. One thing that has remained constant over AI’s 60-year history is its proponents’ high hopes. With enough intelligence, might it be possible to conquer such seemingly impossible problems as poverty and illness—or even win a nuclear war?

The intersection between AI and nuclear warfare became a science fiction cliché more than 50 years ago, but their real-world connections are even older. The earliest AI researchers were deeply involved in national security work and secured government support by suggesting that their theoretical studies would soon translate into practical military applications. Claude Shannon asserted in his foundational 1950 article “Programming a Computer for Playing Chess” that making computers play that venerable game would impart theoretical insights that would make “machines for making strategic decisions in simplified military operations” possible “in the nearterm future” (Shannon, 1950, p. 256). In the mid-1950s, researchers created the earliest working AI programs with support from the U.S. Air Force (Simon and Newell, 1958; Newell, Shaw, and Simon, 1959). Potential applications of such machines soon began appearing in the writings of strategic theorists. In the late 1950s, Herman Kahn postulated the notion of “doomsday machines” that would employ 10 computers programmed to recognize unacceptable enemy provocations and retaliate (Kahn, 1960, pp. 145–154). While Kahn intended these as thought experiments illustrating how not to conduct nuclear strategy, science fiction authors latched onto the idea of intelligent computers controlling nuclear weapons, inspiring numerous novels and such films as Colossus (1970), WarGames (1983), and Terminator (1984).

While fictional thrillers spin tales of nuclear armed computers run amok, real-world attempts to apply AI to nuclear strategic problems tended to be much more mundane. Neither U.S. nor Soviet officials were inclined to entrust launch decisions to computers, both because they jealously reserved this prerogative for themselves and because automating retaliation was not a logical response to difficult strategic problems, such as compellence or crisis stability. The sole notable exception came from the Soviet Union at the end of the Cold War. Perceiving that the United States aspired to a first-strike capability and anxious that they might be the objects of a decapitation strike, Soviet leaders sought measures to ensure that capitalist aggressors would never go unpunished.6 The Soviet Union reportedly considered developing a system that would have automatically launched surviving ICBMs at the United States following a first strike if it could not contact the Soviet political leadership. It seems that the fully automated version, nicknamed the “Dead Hand,” was rejected in favor of a version, dubbed “Perimetr,” that would automatically delegate launch authority to field commanders but would always require a human in the loop (Hoffman, 2009). According to Russian media accounts, the Perimetr system still exists and uses some kind of AI.7 The United States, meanwhile, explored the possibility that AI could be used to bolster its counterforce capability. One late 1980s research project, the Survivable Adaptive Planning Experiment (SAPE), sought to use the AI technology of the time to enable the United States to target the Soviet Union’s mobile ICBM launchers. The SAPE would not control nuclear weapons directly; rather, it would employ expert systems to translate reconnaissance data into nuclear targeting plans that would then be carried out by manned B-2 bombers. The SAPE was just one part of an envisioned suite of systems and capabilities that, if actualized, would have severely challenged the survivability of the Soviet Union’s nuclear arsenal (Roland and Shiman, 2002, p. 305; Long and Green, 2012).

AI and the Emerging Geopolitical Order

Although 20th-century AI struggled to actualize these applications, more-recent advances in computing could release their potential. Such contemporary techniques as deep learning are dramatically advancing machine vision and other signal processing applications, which can enhance autonomy and sensor fusion. Autonomy and sensor fusion may be of paramount strategic relevance because they could greatly improve ISR, ATR, and terminal guidance capabilities. All of these might severely erode the means by which nuclear powers assure the survivability of their nuclear forces. Because increased weapon accuracy has long since undermined the survivability of silo-based ICBMs, the United States, Russia, and China put nuclear weapons on submarines and mobile ICBMs that were deemed more likely to survive a first strike. Technologies that make it more likely that survivable forces (such as submarine and mobile missiles) could be targeted and destroyed make it more plausible that one country might threaten a first strike. This undermines strategic stability, because even if the state possessing these capabilities has no intention of actually using them, the adversary cannot be sure of that. Thus, the capabilities can still be used to pressure potential adversaries and perhaps extract concessions during a crisis. Such a capability does not have to be exploited during a crisis to be politically useful. As Alfred T. Mahan observed, “force is never more operative than when it is known to exist but is not brandished” (Mahan, 1912, p. 105). As long as adversaries fear that the capability may exist, they can be cowed into submission without explicit confrontation—the more powerful state can in effect preemptively “win” the crisis. As a consequence, counterforce targeting capability is an enticing prospect for many despite its potential to compromise strategic stability.

AI technologies could help enable new breakthroughs in tracking and targeting and in antisubmarine warfare or make it easier for high-precision conventional munitions to destroy hardened ICBM silos (Holmes, 2016). Such capabilities would be especially destabilizing because decisionmakers could threaten to employ conventional weapons much more plausibly than any kind of nuclear attack. A conventional threat would place the adversary under enormous pressure during a crisis, which could force it to capitulate—but could also spiral into nuclear war. Such escalation could happen because the adversary felt the need to use its weapons before being disarmed, in retaliation for an unsuccessful disarming strike, or simply because the crisis triggered accidental use.