What happens when artificial intelligence meets modern warfare?
AI Arms Race Escalation
Scenario Analysis
The proliferation of autonomous weapons systems among nation-states has entered a self-reinforcing feedback loop that classical deterrence theory was never designed to handle. As of 2025, over 50 nations have active autonomous weapons programs — a figure that has more than doubled since 2020. The underlying logic is straightforward and inescapable: if your adversary might deploy autonomous systems, you cannot afford not to. This is the trap of the AI arms race.
The structural problem is what RAND analysts have termed the "speed-authority gap." Human decision cycles operate in minutes to hours. Modern AI-enabled targeting and engagement systems operate in milliseconds to seconds. As nations integrate AI more deeply into command-and-control architectures, the window for human deliberation shrinks. This is not merely a tactical concern — it is a systemic transformation of how conflict escalates.
Classical deterrence theory rests on two pillars: the credibility of retaliation and the certainty of attribution. AI arms racing attacks both. Swarm drone attacks, AI-enabled cyber operations, and autonomous loitering munitions can all be designed to obscure attribution, compressing the time between incident and response while muddying the crucial question of who fired first. A Brookings Institution analysis noted that ambiguity of this kind was responsible for the most dangerous near-misses of the Cold War — and the pace of AI proliferation has dramatically increased the conditions for such ambiguity.
Perhaps most concerning is the proliferation cascade dynamic. As China and the United States race to deploy autonomous systems, second-tier powers — India, Turkey, South Korea, Israel, Iran — must respond in kind to maintain regional deterrence. Below them, even non-great powers see autonomous drones as an equalizer, a way to punch above their weight. The result is a global diffusion of lethal autonomous technology across dozens of military organizations with varying safety standards, doctrine, and command accountability.
The misidentification risk compounds this danger. AI targeting systems trained on historical data may misclassify civilian infrastructure, non-combatants, or allied assets. A single high-profile misidentification event — particularly one involving civilian casualties at scale — could trigger escalatory responses before the error is even discovered. RAND's 2024 modeling suggests a 23% probability of an AI-enabled misidentification triggering an escalation sequence involving at least two nuclear-armed states by 2035.
The international governance response has been inadequate. The Convention on Certain Conventional Weapons process has produced no binding treaty on lethal autonomous weapons systems (LAWS). The Bletchley Declaration on AI safety, while historically significant, addressed commercial AI rather than military applications. Without binding multilateral constraints, the arms race will continue on its current trajectory.
Autonomous Weapons + Non-State Actors
Scenario Analysis
This is not a future scenario. It is the present. The commoditization of drone technology, combined with freely available open-source computer vision and machine learning libraries, has placed rudimentary autonomous weapons capability within reach of virtually any organized non-state actor with access to commercial supply chains. The barrier to entry for a functional armed autonomous drone is measured in hundreds of dollars, not millions.
The documented cases are unambiguous. Houthi forces in Yemen have deployed Iranian-supplied Shahed-series loitering munitions to strike Saudi infrastructure at ranges exceeding 1,500 kilometers — demonstrating that even a sub-state actor with state sponsorship can project long-range autonomous strike capability. ISIS conducted over 300 drone operations in Iraq and Syria between 2015 and 2019, including the first documented use of commercially modified quadcopters as precision bombing platforms against Iraqi security forces. Mexican cartels — specifically the CJNG — began deploying surveillance drones with facial recognition systems as early as 2020, later progressing to IED-carrying drones for targeted assassinations.
The proliferation pathways are multiple and difficult to close simultaneously. Commercial technology — DJI Phantom-series and similar consumer drones — is readily available globally and trivially modified. State sponsors, including Iran for the Houthis and Wagner/GRU for certain African militias, provide more capable systems. The dark web provides both finished systems and component packages for AI-enabled guidance payloads. Computer vision models pre-trained on public datasets can be fine-tuned for target recognition with minimal additional data.
The AI component transforms these platforms from remotely piloted nuisances into semi-autonomous threats. A drone equipped with a Jetson Nano or Raspberry Pi running YOLOv8-based object detection can identify and track targets without continuous operator input — defeating GPS jamming, communications disruption, and geofencing countermeasures. The resulting system requires minimal operator skill, operates at ranges that defeat most tactical counter-drone systems, and can be produced in quantity for the cost of a car.
The strategic implications for counterterrorism and domestic security are severe. Current counter-drone (C-UAS) systems are optimized for single or small numbers of threats, carry unit costs that far exceed the drones they destroy, and cannot be deployed at every potential attack vector. A determined non-state actor with $50,000 could deploy a swarm that defeats virtually any point-defense system currently fielded by law enforcement or military forces in the field.
Intelligence community assessments from 2024 note that at least 14 non-state actor groups have active programs to integrate AI-enabled autonomy into their unmanned systems. The 2028 timeline for widespread capable deployment may already be conservative.
AI-Enabled Nuclear Command
Scenario Analysis
On September 26, 1983, Soviet Lieutenant Colonel Stanislav Petrov received an alarm from the Oko early warning satellite system indicating five Minuteman ICBMs had been launched from the United States. The system was confident. Petrov was not. He judged — correctly — that a genuine US first strike would not consist of five missiles, and he did not escalate. He waited. His judgment saved the world. The question this scenario poses is simple: what happens when Petrov is replaced by an algorithm?
Russia's Perimeter system — known in the West as "Dead Hand" — is a semi-automatic nuclear launch system designed to survive a decapitation strike and ensure retaliatory capability. Originally built in the 1980s, Russian strategic communication has strongly implied that Perimeter has been modernized with AI-enabled sensor fusion and autonomous verification capabilities. A system designed to launch without human authorization if it detects certain pre-programmed conditions is, by definition, an autonomous nuclear weapon.
The United States' Nuclear Command, Control, and Communications (NC3) modernization program — funded at over $77 billion through 2035 — explicitly incorporates machine learning for threat identification, sensor fusion, and decision support. The current US posture maintains human authority over the launch decision, but the AI systems feeding that decision are becoming increasingly autonomous in their threat assessments. The gap between "AI decision support" and "AI decision-making" narrows with each system integration.
China's nuclear posture is undergoing its most significant modernization since the 1960s. The PLA Strategic Rocket Force is integrating AI-enabled early warning systems that compress the decision window for Chinese leadership. As China moves from a "no first use" launch posture toward potential launch-under-attack capability, the role of automated detection and decision support becomes critical — and dangerous.
The false alarm risk is the most immediate concern. AI-based early warning systems trained on historical satellite and radar data face adversarial conditions they have never encountered: novel re-entry vehicles, maneuvering warheads, hypersonic glide bodies, and sophisticated decoys. A high-confidence false positive in a crisis environment — when both sides are on elevated alert — could trigger an automated launch sequence before any human has time to intervene.
The existential calculus here differs from all other threat scenarios. A misidentification by an autonomous targeting drone kills people. A misidentification by an AI-enabled nuclear early warning system during a crisis could end civilization. The margin for error is not percentage points — it is zero.
Drone Swarms vs Aircraft Carriers
Scenario Analysis
The USS Gerald R. Ford (CVN-78) cost $13.3 billion to build and carries approximately 5,000 personnel. An adversary capable of fielding 10,000 AI-guided drones at $500 each — a total investment of $5 million — has created a cost-exchange ratio so asymmetric that it fundamentally challenges the viability of carrier-centered naval power projection. This is not a hypothetical. It is the precise strategic problem that the Center for Strategic and Budgetary Assessments identified in its landmark 2015 analysis, updated repeatedly as drone technology has matured.
The saturation attack problem is mathematical. Current carrier strike group point defense — the Phalanx CIWS, Evolved Sea Sparrow, SM-6 — is designed to defeat a limited number of simultaneous threats. Phalanx fires at 4,500 rounds per minute against a theoretical one-target-at-a-time engagement. Against 100 simultaneous autonomous drones arriving from multiple vectors, the math does not close. The fleet's Aegis Combat System can engage multiple targets simultaneously, but each Standard Missile round costs $3–4 million. Shooting a $4M missile at a $500 drone is an exchange ratio that bankrupts even the United States.
The Chinese PLA Navy and PLA Rocket Force have explicitly oriented their anti-access/area-denial (A2/AD) strategy around this asymmetry. The DF-21D and DF-26 carrier-killer ballistic missiles — deployed since 2015 and 2018 respectively — are complemented by sea-launched and air-launched autonomous systems, submarine-delivered drone swarms, and shore-based systems that collectively create a multi-domain saturation envelope. In any Taiwan Strait contingency, a US carrier strike group would face simultaneous threats from ballistic missiles, hypersonic glide vehicles, submarine-launched torpedoes, and autonomous drone swarms operating in coordination.
The CSBA's 2023 analysis of carrier vulnerability concludes that within the DF-26's 4,000km range — which covers virtually the entire Western Pacific — a carrier strike group cannot be confident of operating without catastrophic losses. The analysis recommends shifting to distributed maritime operations with smaller, more numerous surface combatants and submarine-centric power projection. The 2022 National Defense Strategy implicitly acknowledged this by reducing the F-35C carrier variant procurement while increasing submarine funding.
The counter-arguments are real and important. Directed-energy weapons (DEW) — the Navy's AN/SEQ-3 Laser Weapon System and next-generation High Energy Laser systems — offer a different cost exchange: a $1 laser shot versus a $500 drone. Electromagnetic pulse systems and advanced electronic warfare can disrupt drone swarm communications and guidance. Carrier air wings retain unique strike capability at ranges beyond drone systems. And the political-deterrent value of a carrier battle group remains high independent of its raw military vulnerability.
Nevertheless, the trajectory is clear. The era of the aircraft carrier as an unchallenged symbol of sea control power is ending. The question is not whether drone swarms will challenge carrier dominance, but when — and whether US naval doctrine will adapt before the Taiwan Strait becomes the test case.