Interview
Will AI End Mutual Assured Destruction?
- Nuclear deterrence's core mechanism: In a nuclear world, the outcome of bargaining contests is determined less by the balance of material power (economy, resources, weapon counts) and more by the "imbalance of nerves"—specifically, which side is more resolved, willing to accept higher escalation risks, or cares more about specific outcomes.
- The secure second strike: The central pillar of deterrence is a state's ability to absorb a first strike and still retaliate with devastating force; if this capability is compromised, a state could theoretically dominate adversaries without fear of reprisal.
- Three pathways to undermining deterrence: AI could potentially erode secure second strike capabilities by enabling:
- Splendid first strikes: Using AI to pinpoint and destroy every nuclear asset (including submarines and mobile launchers) simultaneously.
- Disabling Command, Control, and Communications (NC3): Using AI to penetrate cyber networks, disabling the systems required to authorize and execute a retaliatory strike.
- Overwhelming missile defense: Using AI to improve interceptors, track trajectories, and distinguish decoys from warheads to neutralize a rival's ability to threaten retaliation.
- Nuclear submarine survivability: Ballistic missile submarines (SSBNs) are currently considered the most survivable leg of the nuclear triad due to:
- The vast, noisy, and optically opaque volume of the ocean.
- Advanced acoustic silence engineering making detection extremely difficult (submarines sometimes collide because they cannot hear each other).
- AI's potential role: ML could fuse data from sonar, magnetic anomaly detectors, and satellite radar to track faint signatures, but technical limits on underwater physics and the need for millions of sensors make "transparency" of the ocean improbable.
- Countermeasures: States can deploy decoys, jam signals, and utilize noisy commercial shipping lanes to maintain uncertainty, meaning the US Navy likely cannot achieve 100% tracking confidence.
- Road mobile missile survivability:
- AI's potential role: AI could dramatically speed up the processing of satellite and intelligence data to identify mobile launchers hidden in vast territories, potentially shrinking the search area enough to enable a first strike.
- Countermeasures: Low-tech solutions like camouflage netting, decoys, and anti-satellite weapons (to blind sensors) remain highly effective and do not require AI parity to maintain survivability.
- Historical context: During the Cold War, the US was more effective at tracking Soviet subs due to geography (choke points) and Soviet submarine noise, suggesting "windows of vulnerability" can exist, though not enough to guarantee a successful first strike.
- Missile defense limitations:
- Technical constraints: Defense systems must detect, track, and destroy hundreds of high-speed targets within ~30 minutes, a task likened to "shooting a bullet with a bullet."
- Decoy economics: It is cheaper to build decoys than interceptors, allowing attackers to overwhelm defenses economically.
- AI's impact: AI may improve trajectory prediction and decoy discrimination, but adversaries can poison training data, and the sheer physics of high-speed intercepts limit reliability against large-scale barrages.
- Alternative delivery: Even with perfect defense, states can use exotic delivery methods like uncrewed undersea vehicles (e.g., Russia's Poseidon) or smuggled pre-positioned warheads.
- NC3 (Command, Control, Communications) resilience:
- Physical hardening: Deep bunkers (700m underground) and airborne command posts are difficult to destroy even with improved targeting.
- Cyber vulnerabilities: AI could enhance cyber attacks to delay response times or disrupt early warning systems, but redundancy, analog backups, and the difficulty of testing malware against live nuclear networks create significant barriers.
- Automatic retaliation: Systems like Russia's "Dead Hand" (Perimeter) and UK submarine "Letters of Last Resort" ensure retaliation can occur even if the national command authority is decapitated.
- Fast takeoff scenarios:
- Speed of adaptation: The critical risk factor is whether AI progress outpaces an adversary's ability to adapt defensively.
- Invisible breakthroughs: AI-enabled intelligence processing could create sudden, invisible windows of vulnerability where one state gains confidence it has 100% tracking of adversary assets before the other side detects the change.
- Political lag: Even if technology advances overnight, political, institutional, and doctrinal changes take much longer, potentially leaving a dangerous gap where states might consider a first strike despite the uncertainty.
- Decision-making uncertainty:
- Trust in AI: A president would likely hesitate to accept AI-generated claims of 100% certainty regarding adversary nuclear locations, given the system's lack of testing and potential for deception.
- Risk of failure: Launching a first strike requires near-perfect certainty; the existential risk of a failed attempt (leaving the adversary with a surviving second strike) creates a massive disincentive to act, even if a "splendid first strike" seems technically feasible.
- Historical precedent: The US possessed nuclear monopoly for years without launching a first strike, demonstrating that technological advantage does not automatically translate to political dominance or willingness to use weapons.
- Non-nuclear coercion limits: Even if AI turbocharges economic or cyber capabilities, it may not allow a state to coerce nuclear-armed adversaries (e.g., US/South Korea vs. North Korea) on core interests like regime survival, as nuclear deterrence remains a binding constraint on political coercion.
- Forward-looking recommendations for governments:
- Cross-disciplinary dialogue: Establish formal channels between AI experts (who understand emerging tech) and nuclear experts (who understand deterrence dynamics).
- Vulnerability assessments: Conduct rigorous reviews of nuclear systems, particularly cyber vulnerabilities, to ensure they are resilient against AI-driven attacks.
- Cautionary rhetoric: Avoid public statements that frame AI as a "wonder weapon" to prevent exacerbating arms races and security dilemmas.
- Maintain communication: Strengthen arms control dialogues and crisis communication pathways to prevent inadvertent escalation during periods of rapid AI development.
- Monitor "windows": Actively monitor for periods where AI advancements might outpace defensive adaptations, as these windows create the highest risk for instability.