The conversation on how artificial intelligence (AI) could impact warfare is still very young. In the past few years, much of this dialogue has revolved around the debate on Lethal Autonomous Weapons Systems (LAWS — also known as killer robots), which currently takes place within the framework of the UN Convention on Certain Conventional Weapons (CCW). As its name suggests, the CCW focuses on conventional weapons, so it has a major blind spot on issues related to nuclear weapons and strategic stability. The transformative potential of AI, however, is also relevant for nuclear weapons and doctrines. AI could even be a driver of great ‘entanglement’ between the two areas. This convergence of AI and nuclear weapons deserves greater scrutiny.
The connection between AI and nuclear weaponry is not new. In fact, AI has been part of the nuclear deterrence architecture for decades. As early as in the 1960s, the United States and the Soviet Union saw that the nascent field of AI could play a role in the development and maintenance of their retaliatory capability — that is, the capability to respond to a nuclear attack, even by surprise.
They pursued the development of AI systems that could make their command and control process more agile and give decision-makers more time to focus on what really mattered: deciding whether to launch a nuclear strike or not. Early application of AI included automating threat detection, logistical planning for the transmission of launch orders, and missile targeting and guidance.
Early on, nuclear-armed states not only identified the appeal of AI for nuclear deterrence, but they also saw its limitations. Given the dramatic consequences that a system failure would have, they were reluctant to hand over higher-order assessments and launch decisions to AI systems: a human had to remain ‘in the loop’. The Soviet Union is the only country that pursued the development of fully automated command and control systems for nuclear weapons. This system, known as the Dead Hand, was however meant to be activated only in the exceptional case of a decapitating attack on the Soviet nuclear command and control.
What might change with the current AI renaissance, which is seeing breakthroughs in the areas of machine learning and autonomous systems? Recent advances in AI could be leveraged in all aspects of the nuclear enterprise.
Machine learning could boost the detection capabilities of extant early warning systems and improve the possibility for human analysts to do a cross-analysis of intelligence, surveillance, and reconnaissance data. Machine learning could be used to enhance the protection of the command and control architecture against cyberattacks and improve the way resources, including human forces, are managed. Machine learning advances could boost the capabilities of non-nuclear means of deterrence: be it conventional (air defence systems), electronic (jamming) or cyber.
Autonomous systems could be used to conduct remote sensing operations in areas that previously were hardly accessible for manned and remotely-controlled systems, such as in the deep sea. Autonomous unmanned systems such as aerial drones or unmanned underwater vehicles could also be seen by nuclear weapon states as an alternative to intercontinental ballistic missiles (ICBMs) as well as manned bomber and submarines for nuclear weapon delivery.
These would be recoverable (unlike missiles and torpedoes) and could be deployed in ultra-long loitering periods — days, months or even years. At least one nuclear-armed state is already considering that possibility: In 2015, Russia revealed that it was pursuing the development of a nuclear-armed unmanned submarine, called Status-6.
Will the adoption of such systems fundamentally transform the field of nuclear strategy? The answer is no, at least not in the near term, for three reasons.
First, these technologies reinforce rather than fundamentally alter the existing application of AI in nuclear force-related systems.
Second, the field of nuclear weapon technology is renowned for its conservativeness; it has been historically slow at integrating new technologies. The US military for instance allegedly still uses 8-inch floppy disks to coordinate nuclear force operations. In that regard, machine learning and autonomous systems have some critical technical limitations that would make a rapid adoption unlikely in the near future.
Machine learning systems operate like black boxes, which makes them potentially unpredictable, while the reliability of advanced autonomous systems is also technically hard to establish. Nuclear-armed states would have to crack difficult testing issues associated with the design of these systems to be confident that they can be used in a predictable and reliable manner and be certified for use.
Third, the technology is not at the stage where it would allow nuclear-armed states to credibly threaten the survivability of each other’s nuclear second-strike capability. Some experts have argued that a large-scale deployment of autonomous unmanned systems for remote sensing could make the continuous at-sea deterrence obsolete. In light of the current stage and development trajectory of AI technology and other key enabling technologies (such as sensor and power technology), this is bound to remain a very theoretical scenario for the foreseeable future.
If recent advances in AI are unlikely to completely undermine the foundation of nuclear strategy, they could, without a doubt, have both a positive and negative impact on strategic stability.
On the one hand, recent advances of machine learning and autonomous systems could enhance stability as they provide nuclear weapon states with better information and better decision-making tools for time critical situations, which would reduce the risk of miscalculation and accidental escalation. They could moreover generate new possibilities for the arms control community to monitor nuclear weapon-related developments and conduct verification operations.
On the other, the adoption — or even suspected adoption — of new AI capabilities by one or several nuclear-armed states could also incentivise other states — be they nuclear-armed or not — to respond with destabilising measures that could increase the likelihood of a nuclear conflict. This could include entering into an arms race, doubling down on the modernisation of nuclear arsenals, renouncing a ‘no first use’ policy, increasing alert status, or further automating nuclear launch policies. Historical events like the 1983 Petrov incident (where the Soviet early warning systems wrongly detected a US nuclear attack) have also shown that AI technology could be the cause of an accidental or inadvertent escalation into a nuclear conflict.
The risks posed by the convergence of AI and nuclear weapon technology are not necessarily new. Some of them have been known for years. This means that solutions to address them may already exist. There might be no need to reinvent the wheel. ‘No first use’ policies, a commitment to lower the alert status of nuclear arsenals, as well as more openness about nuclear modernisation plans and information sharing via different dialogue tracks are measures that could clearly help to start mitigating the destabilising potential of nuclear-related AI applications. It is essential that nuclear-armed states take note of the importance of this issue as part of the bilateral and multilateral talks on nuclear risk reduction.
•••
Dr Vincent Boulanin wrote this article as a contributor to AI & Global Governance, an inclusive platform for researchers, policy actors, corporate and thought leaders to explore the global policy challenges raised by artificial intelligence.
The opinions expressed in this article are those of the author and do not necessarily reflect the opinions of the United Nations University.
