Interview
Will AI End Mutual Assured Destruction?
- As long as nuclear deterrents remain functional, AI-driven economic or military advantages are unlikely to enable states to completely dominate nuclear-armed adversaries, with the persistence or erosion of these deterrents in the coming AGI era fundamentally shaping the risks and stakes of AI competition.
- States may continue to temper actions due to fear of nuclear reprisal, creating significant limits on imposing political preferences against major nuclear-armed powers, though specific windows of vulnerability could emerge if AI undermines second-strike capabilities.
- AI technologies may advance capabilities in tracking nuclear assets through machine learning integration of thousands of sensors, fusing noisy data from sonar and magnetic anomaly detectors to identify small submarine wake patterns via satellite radar, and coordinating autonomous underwater vehicles to build persistent surveillance nets.
- Operational challenges to AI-driven submarine tracking include increasing ocean noise from commercial shipping, the enormous number of uncrewed vehicles required for ocean transparency, and the difficulty of deploying millions of sensors in contested waters due to covert communication limits and sabotage risks.
- States are likely to employ countermeasures such as signal jamming, deceptive decoys, and acoustic playback, while underwater sensor networks in regions close to home territories may protect submarines, keeping the detection and targeting of mobile forces largely probabilistic.
- Ballistic missile submarines are projected to remain reliable second-strike forces over the next 20 years and beyond, although windows of confidence in tracking assets may emerge similar to Cold War dynamics, and AI may reduce search areas for road-mobile launchers using open-source image training.
- Defenders may adopt low-tech solutions like road netting or decoys, and attackers may utilize anti-satellite weapons to create coverage gaps, while a measure-countermeasure cycle could potentially increase nuclear domain instability.
- AI may improve missile defense by predicting trajectories and distinguishing warheads from decoys, though significant limits persist due to the need for large, reliable datasets and the ability of adversaries to manipulate missile tests to poison machine learning models.
- Missile defense architectures require years to develop, allowing adversaries time to adapt, while attackers may maintain advantages via uncrewed undersea vehicles or smuggling, with Russia reportedly developing nuclear-powered autonomous torpedoes like Poseidon.
- Sophisticated AI-enhanced cyber operations could penetrate command and control networks, disable early warning systems, or target satellites, though persistent access risks detection and retaliation, and states may expend significant resources on redundant, diverse NC3 systems.
- Move-countermove dynamics may increase the risk of accidents or speed up adaptation, but there is no guarantee states will adapt at the same speed, with nuclear-armed states expected to prioritize the survivability of their second-strike forces given the existential stakes.
- Debate exists regarding damage limitation strategies, where one view holds that only 100% confidence in destroying all adversary weapons matters, while another suggests AI can meaningfully limit damage even without total success, making the definition of meaningful damage a political rather than purely technological question.
- Ethical, political, and international norms, such as public intolerance for preemptive strikes or historical precedents of technology not translating to political dominance, may constrain the use of full power, though fast AI takeoffs could exacerbate instability if they outpace state adaptation.
- Governments may need to increase state capacity by integrating AI and nuclear experts, conducting rigorous reviews of nuclear systems for AI-exploitable cyber vulnerabilities, and calibrating race rhetoric to avoid exacerbating competition.
- Policy recommendations include strengthening arms control dialogues, reinforcing ethical and legal constraints on nuclear use, and recognizing that economic or cyber channels are unlikely to coerce nuclear peers while second-strike capabilities remain intact.