With nine countries now possessing nuclear weapons, arms control agreements collapsing, and geopolitical competition intensifying, this article explores the dangerous potential for artificial intelligence to automate nuclear launch decisions. It discusses how AI could make nuclear threats more credible, drawing parallels to Cold War systems, and examines the inherent risks and policy implications for international security and deterrence.
Why Nuclear Coercion Is Challenging
Nuclear coercion, using threats to achieve political goals without force, typically faces two main constraints. Firstly, mutual assured destruction (MAD) between nuclear-armed states makes carrying out a nuclear threat illogical due to the certainty of devastating retaliation, thus undermining credibility. Secondly, the 'nuclear taboo' or non-use norm implies high political, economic, and social costs for employing nuclear weapons, further reducing the credibility of such threats.
How Automation Could Change the Game
Automating nuclear launch systems could enhance the credibility of nuclear threats by removing human elements like rationality and the capacity for de-escalation. While historical precedents like the Soviet 'Dead Hand' system existed, modern AI introduces new and significant risks, including malfunctions, hacking, and automation bias (humans overly trusting computers). Despite these dangers, countries might be tempted to automate nuclear command and control to prevent decapitation, respond to rapid conventional defeat, or leverage conventional disadvantages to achieve strategic objectives.
Nuclear Automation “Works” (To Some Extent)
A survey experiment involving U.K. parliament members and the general public demonstrated that automated nuclear threats, in a hypothetical scenario of a Russian invasion of Estonia, were perceived as more credible. Respondents were more willing to back down when Russia's threat was automated, suggesting that automation can meaningfully influence perceptions of threat. However, the study notes challenges in verifying such systems and leaders' inherent reluctance to relinquish control over crucial decisions, though high stakes or autocratic regimes might make such claims more believable.
Why Nuclear Automation Is Different
Given the coercive advantages of automated nuclear systems, states must anticipate their adversaries potentially adopting them. Countermeasures include publicly declaring consequences for developing these systems and bolstering second-strike capabilities to reduce incentives for defensive automation. Multilateral agreements prohibiting automated nuclear launch, though hard to verify, could formalize non-compliance costs. NATO should prepare through wargaming to respond to automated nuclear threats, while avoiding putting Armageddon on autopilot themselves due to the unacceptable risks of malfunction, hacking, and escalation inherent in such systems.
