Why Cold War Arms Control Breaks Down Under AI Risk Dynamics

Why Cold War Arms Control Breaks Down Under AI Risk Dynamics

Cold War arms control depended on counting physical objects in space or on the ground; artificial intelligence risk governance fails when negotiators treat software weights and compute clusters like nuclear warheads. Calling for Washington and Beijing to replicate mid-twentieth-century treaties mistakes a thermodynamic and institutional regime built around scarce, observable hardware for a dual-use digital capability that scales through code, commercial supply chains, and diffuse deployment. This analysis maps the core logic behind arms-control analogies for artificial intelligence, separates verified mechanics from speculative projections, and establishes a structural framework for managing bilateral AI risk without relying on brittle treaty templates.

The central thesis of this analysis is simple: treaty mechanics designed for verifiable physical stockpiles cannot govern general-purpose AI systems without first solving observability, attribution, and incentive alignment. To understand where the analogy helps and where it distorts strategic reality, we must examine how historical deterrence worked, why AI breaks those assumptions, and what practical guardrails remain feasible between the United States and China. Expanding on this topic, you can find more in: Hardening Data Centres Against Drone Warfare Why Standard Perimeter Security Fails.

Deconstructing the Cold War Analogy

Proposals to apply Cold War arms control to artificial intelligence usually rest on a familiar chain of reasoning. Advocates observe that Washington and Moscow faced existential security dilemmas, volatile crisis dynamics, and deep mutual distrust, yet still established communication channels, crisis hotlines, confidence-building measures, and numerical or deployment limits to reduce the probability of catastrophic escalation. Transposed onto US-China AI competition, the argument suggests that policymakers can stabilize rivalry through agreed red lines, bilateral technical dialogues, and verifiable ceilings on frontier model training or military deployment.

That logic contains a real insight alongside a major structural flaw. The valid insight is process-oriented: crisis stability requires predictable signaling and open channels before an incident occurs. During the Cold War, direct communication and agreed norms helped prevent miscalculation when radars glitched, patrols drifted, or proxy crises flared. Applied to AI, avoiding unforced escalation over autonomous weapons, cyber-enabled infrastructure disruption, or misread military posture is a genuine strategic need. Observers at MIT Technology Review have provided expertise on this matter.

The structural flaw lies in treating treaty verification and force ceilings as transferable mechanisms. Cold War treaties such as SALT I, the ABM Treaty, and New START succeeded—where they did succeed—because strategic delivery systems had distinct physical signatures. Intercontinental ballistic missiles required large launch complexes, test flights, telemetry, and specialized industrial facilities monitored by national technical means such as reconnaissance satellites and seismic sensors. An ICBM silo could not be quietly replicated inside an ordinary office building or compiled from an open-source repository overnight.

Artificial intelligence operates under a fundamentally different cost and observability function. A frontier model is not a single physical artifact; it is a combination of curated data pipelines, training code, model weights, post-training alignment runs, and inference infrastructure. Model weights can be copied across storage media, compressed, fine-tuned, or exfiltrated without leaving a satellite-visible signature. Even when training relies on large clusters of specialized chips, the boundary between civilian cloud commerce, research experimentation, and military-relevant capability is porous.

Known Facts vs. Educated Hypotheses in AI Governance

Clear strategic analysis requires separating established facts from educated hypotheses.

Known facts include the dual-use nature of modern machine learning architectures, the global distribution of AI research talent and cloud infrastructure, and the absence of an international verification regime capable of auditing private or state-controlled training runs with high confidence. It is also a known fact that historical arms control treaties took years of negotiation, required shared baseline definitions of covered systems, and still experienced disputes over definitions, modernization loopholes, and compliance.

Educated hypotheses include claims that frontier AI systems will inevitably trigger sudden strategic instability, that bilateral compute caps can be reliably enforced across jurisdictions, or that technical verification tools—such as hardware telemetry, cryptographic watermarking, or third-party audits—will mature fast enough to underpin binding treaties. Treating hypotheses as settled facts leads negotiators to build fragile agreements around assumptions that adversaries can easily circumvent.

The Structural Mismatch of AI Risk

To evaluate why simple treaty borrowing falls short, we can examine four structural properties that separate AI risk from nuclear arms control.

  • Observability asymmetry: Nuclear delivery vehicles are large, capital-intensive, and difficult to conceal in operational numbers; AI model weights and algorithmic improvements are compact, intangible, and easy to copy or modify quietly.
  • Dual-use entanglement: Nuclear weapons have no substantial everyday civilian commercial substitute; frontier AI models and high-performance compute clusters power search, enterprise software, scientific modeling, and military logistics simultaneously.
  • Pace of technical change: Strategic nuclear delivery technology evolved over decades with long procurement cycles; model architectures, training techniques, and post-training methods shift in months, outrunning static treaty definitions.
  • Attribution ambiguity: A nuclear launch carries clear physical and geopolitical signatures; an AI-enabled cyber intrusion, disinformation campaign, or autonomous system malfunction can obscure actor intent, supply-chain origin, or whether human operators or model behaviors drove the outcome.

Why Deterrence Math Changes When Systems Are Software

Classical nuclear deterrence relied heavily on mutual vulnerability and second-strike survivability. Each superpower knew what a catastrophic exchange looked like and understood that retaliatory force could survive a first strike. That clarity created a sobering threshold: crossing from conventional or proxy competition into nuclear use carried immediate, unmistakable devastation.

AI risks rarely present a single, unmistakable threshold. Instead of a sharp cliff, AI risk behaves more like a gradient of compounding failures and escalatory pressures. A military command system might rely on AI-assisted target prioritization or intelligence synthesis that subtly biases human decisions under time pressure. An autonomous drone swarm or cyber agent might behave unpredictably in novel environments without any leader intending a strategic crisis. Because the path to severe harm can run through gradual erosion of reliability, automated speed, or ambiguous incidents, classic deterrence signaling struggles to map cleanly onto the problem.

The second limitation of the Cold War template is domestic political economy. In both the United States and China, commercial technology firms, academic labs, state-backed enterprises, and defense institutions are deeply interwoven. Unlike a narrow nuclear weapons complex guarded by a small cadre of specialists, AI development spans massive economic ecosystems where commercial competitiveness is tied directly to national prestige and industrial policy. A government cannot easily freeze a fast-moving commercial technology sector without risking economic stagnation relative to rivals—unless every major actor accepts strict, verifiable constraints, which remains far from reality.

Three Pillars of Practical Bilateral Risk Reduction

Rejecting a literal Cold War treaty does not mean accepting unmanaged risk. Instead of chasing comprehensive arms-control pacts that neither side can verify, Washington and Beijing can focus on three narrower, mechanism-driven pillars where interests partially overlap.

Pillar 1: Crisis Communication and Escalation Guardrails

The highest-value lesson from the Cold War is process discipline during high-tempo crises. Both capitals benefit from clear military-to-military and technical channels that reduce the chance of misinterpreting an automated alert, an uncrewed platform incident, or an AI-accelerated cyber event as a deliberate strategic attack.

Operationalizing this pillar does not require revealing proprietary algorithms or sensitive defense architectures. It requires agreed protocols for crisis hotlines, clear distinctions between routine testing and operational deployment in sensitive theaters, and shared recognition that keeping human command authority visible in high-consequence decisions lowers accidental escalation risk.

Pillar 2: Narrowly Defined Red Lines Around High-Consequence Domains

Broad bans on "dangerous AI" fail because definitions slip as capabilities evolve. Narrow, domain-specific guardrails hold up better over time. Both sides have a pragmatic interest in avoiding loss of human control over nuclear command-and-control systems, preventing uncontained biological risks, and limiting reckless automation in critical infrastructure where cascading failures can rebound onto the initiator.

Even here, enforcement relies on unilateral safeguards, export and supply-chain controls, and domestic accountability rather than trusting an adversary's word alone. Each side should build internal verification, red-teaming, and containment standards for its own systems while monitoring external indicators of risky deployment.

Pillar 3: Transparency on Incidents and Shared Technical Baselines

Trust is scarce in strategic competition, so transparency works best when tied to shared hazard awareness rather than goodwill. Sharing lessons from near-misses, model failure modes, evaluation methodologies, and incident-response practices helps prevent common blind spots.

Consider an illustrative conceptual example: two rival airlines competing fiercely on routes and ticket prices still share flight-safety data, incident reports, and air-traffic rules because a mid-air collision destroys both businesses. Strategic rivals are not airlines, and national security overrides commercial convenience, yet the underlying logic holds where shared catastrophic risk exists. If an AI system exhibits unexpected deceptive behavior, tool-use errors, or safety bypasses during evaluation, understanding those failure modes improves defensive readiness on both sides without surrendering competitive advantage.

Limitations and Failure Modes of Risk Frameworks

Every governance approach carries failure modes, and pretending otherwise weakens strategy.

  1. Verification evasion: Any agreement that depends on self-reporting of model capabilities or training runs faces strong incentives to understate performance, split training across smaller clusters, or shift work into covert facilities.
  2. Definition capture: Negotiators who define risk solely by parameter counts or compute thresholds risk regulating yesterday's bottleneck while new algorithmic efficiencies or agentic workflows bypass the rule.
  3. Asymmetric compliance: If one side enforces strict internal safety and deployment limits while a rival treats constraints as optional, the compliant side may face a perceived capability gap unless it pairs guardrails with resilient technical defenses.
  4. Performative diplomacy: High-level dialogues can create an illusion of stability while underlying military integration, dual-use procurement, and autonomous systems development accelerate unchecked.

A common question in public debate is whether binding international treaties are still possible in the long run. They remain theoretically possible only if verification technologies advance dramatically and states converge on narrow, high-consequence prohibitions—much like treaties governing specific chemical or biological agents—rather than sweeping caps on dual-use computation. Until those conditions exist, treating non-binding guardrails as binding guarantees is a strategic mistake.

The Strategic Play for Decision-Makers

Stop pitching Cold War arms control as a turnkey blueprint for US-China AI governance, and build a layered risk-management architecture anchored in observable bottlenecks, crisis stability channels, and domestic resilience. Focus bilateral engagement on narrow, high-consequence failure modes—especially nuclear command automation, critical infrastructure escalation, and crisis communications—while investing internally in hardware supply-chain visibility, rigorous pre-deployment evaluations, and rapid incident containment. Treat every diplomatic commitment as a supplement to verifiable defenses, never a substitute for them, so policy remains grounded in what can actually be observed and enforced rather than historical nostalgia.

MJ

Matthew Jones

Matthew Jones is an award-winning writer whose work has appeared in leading publications. Specializes in data-driven journalism and investigative reporting.