Military Artificial Intelligence and the Future of Warfare in a Multipolar World

The "Oppenheimer Moment" for AI

Over the past quarter century, the deployment of artificial intelligence in military affairs has undergone substantial qualitative change, producing what analysts have termed an "Oppenheimer Moment"[1]: a critical juncture in which previously stable paradigms are overturned by the non-linear advance of disruptive technologies, generating ethical dilemmas and security risks that resist easy solution. The transformations in the military use of AI under way since 2024 are an apt analogue to the Manhattan Project, which produced history's first atomic device. The contemporary observer thus bears witness to a transformation in the organizing principles of war and warfighting — principles that, barring catastrophe, will prevail throughout this century, one in which the collapse of unipolar U.S. hegemony is triggering contests for hierarchical repositioning across a range of geopolitical chessboards.

Humans almost off-the-loop

From Turing's breaking of Nazi codes to Wiener's principles of cybernetics, through the first neural-network computer built by Rosenblatt in 1956, to the U.S. Strategic Computing Initiative of 1983, computer engineering was long regarded as a supporting element to intelligence, decision-making, and military action conducted by human beings. In the United States, the 2014 Third Offset Strategy and the 2018 National Defense Strategy — along with their Chinese and Russian equivalents — preserved this orientation: accelerating the integration of big data and machine learning into C4ISR (Command, Control, Communications, Computers, Intelligence, Surveillance and Reconnaissance) workflows over which human operators retained full control. The goal was to expand situational awareness on the battlefield and with respect to force status, while keeping human agents predominantly in-the-loop or on-the-loop — sovereign cognitive and decision-making agents, supervising and retaining the ability to interrupt automated systems' engagement. On-the-loop systems had already existed for decades within specific point-defense niches (Phalanx CIWS, Aegis, Patriot), but their generalization to networked offensive C4ISR remained limited. The prospect of off-the-loop military systems — cognition and decision-making entirely artificial, without real-time human oversight — remained remote, largely a matter of speculative discussion.

This picture has changed dramatically, in an extremely short span of time, since 2023, driven by the systemic ripple effects of geopolitical tensions in the Middle East and Eastern Europe. Artificial intelligence models have emerged as the spearhead of military computing, moving from an ancillary role to become the backbone of the strategic infrastructure of the U.S. and Chinese armed forces, and an instrument of profound tactical importance for Russia. Far beyond intelligence and surveillance operations, targeting systems have raised the speed of combat to previously unimaginable levels, while data platforms now manage the logistics of attack and defense across thousands of objects, monitoring their parameters and suggesting courses of action in real time. This qualitative leap compresses the decision time between detection and targeting to a scale human cognition can no longer keep pace with, reducing human intervention to a residual act of ratification. The result is a de facto off-the-loop regime, even as doctrine insists on preserving, de jure, a human presence in the decision chain.

Thus, in the third decade of the twenty-first century, algorithmic warfare has emerged.

Competing Military AI Doctrines

The period from 2020 to 2026 has come to represent the moment at which the great powers formally recognized "decision dominance" as the principal capability to be achieved by their defense-industrial bases — bases in which major technology companies are now inevitably embedded. What is thus being acknowledged is the new status of artificial intelligence as one of the foremost instruments of deterrence in the international arena. The technological maturation that produced substantial advances in large language models (LLMs), and in other platforms, had the expected consequence of reorienting the strategic doctrines of the United States, China, and Russia. This reorientation, however, has not produced convergent doctrinal frameworks, since it has been strongly shaped by particular geopolitical determinants bearing on decisions made in Washington, Beijing, and Moscow.

America's "AI-First"

The U.S. government, reacting to the progressive erosion of its unipolar hegemony, adopted in January 2026 a posture of maximum readiness for large-scale conflicts against enemies of comparable technological standing — revisionist powers prepared to expand their margins of freedom within the interstate system. The memorandum "Artificial Intelligence Strategy for the Department of War"[2] established that the country's structural capabilities — a world-leading innovation ecosystem, technology industry, capital markets, and operational data accumulated over decades — should be marshaled to guarantee the armed forces absolute supremacy in decision-making systems ("AI Overmatch"). The government's willingness to subordinate the American innovation system to national strategic interests is evident in its rejection of any limits imposed by technology companies' codes of ethics; suppliers have been contractually required to grant the government "any lawful uses" of their products, particularly following the clash between the Department of War and Anthropic in March 2026, arising from the company's complaint that its Claude Acceptable Use Policy — which bars Claude's use in autonomous weapons systems — had been violated.

The memorandum thus reinforced a doctrinal dimension privileging the strategic level over an exclusively tactical focus. In its pursuit of decision superiority, Washington's objective is to outpace every opponent in the speed with which sensor data is processed, intelligence is synthesized, and the kill chain is closed. Platforms such as GenAI.mil enable the integration of frontier generative models (Gemini for Government, Google Cloud) into classified and unclassified networks, granting millions of service members and civilian employees access to AI tools for tasks ranging from logistics management to detection, target identification, and lethal strike, including drone-swarm deployment.

D-DIL environments and the conflict in Ukraine

At the tactical level, U.S. AI doctrine has been tested both in simulated conflicts and on real battlefields — through Ukrainian forces, serving as a NATO proxy against Russia, and through Operation Epic Fury, launched this past February against Iran's military forces. Intelligent systems associated with LLMs have been deployed for target detection, real-time intelligence analysis, and "what-if" simulations, assisting commanders in selecting missile and stealth-drone strike options. The overarching thrust of U.S. AI strategy is to develop every capability within reach to achieve maximum compression of the OODA loop (Observe, Orient, Decide, Act): converting raw information into kinetic action in the shortest possible time, so as to overwhelm any conventional adversary response before it can attain reasonable situational awareness.

Concern is nonetheless mounting over the vulnerability of these highly centralized systems in Denied, Degraded, Intermittent, and Limited (D-DIL) communications environments, where electromagnetic-warfare actions can block high-bandwidth contact between field assets and command-and-control centers. While Project Linchpin channels efforts toward "Edge AI" — processed at the tactical edge, with greater decentralization and reduced vulnerability to EW measures — CENTCOM operations in Iran, together with Palantir's experience in Ukraine, have already yielded proofs of concept for AI use in degraded environments. The Saker Scout and Hunter drones (built in Ukraine with Western embedded technology) can autonomously navigate and identify sixty-four different target types. This data is transmitted by radio frequency to the Delta system (operated with Palantir software), where it is aggregated and processed to direct guided munitions. In the event of jamming, the drone retains its data and transmits it in a burst once connectivity is restored.

Palantir, Maven Smart System

Palantir Technologies' engagement in providing Ukrainian forces with network-centric warfare capability has been consolidating Eastern Europe as a testing ground for what may amount to a revolution in military affairs. The deployment of the Maven Smart System (MSS) has prolonged Ukraine's war effort, sharply reducing troops required per unit of terrain: positions once demanding two battalions per kilometer are now held by minimal squads in small shelters, supported by swarms of surveillance and strike drones and a lethality-optimization system that identifies targets and recommends the most effective precision-strike methods.

The American MSS sustains a multi-source data-fusion regime; its interface with LLMs generates a graphical, real-time picture of the target network, risks, and the disposition of adversary and allied assets, allowing command agents to receive, in plain language, guidance on optimized courses of action aimed at maximizing lethality with maximum speed — including munition options, attack vectors, and legal justifications — while granting these agents the ability to activate battlefield units and assets through simple commands.

Maven further ensures interoperability between Ukrainian forces and NATO member states, such as the Royal Marines and the Polish Volunteer Corps, and enables integration with France's Safran.AI for satellite-imagery exploitation and Britain's Hadean system for generative COA (course of action) simulations. It has proven central to the NSATU mission (NATO Security Assistance and Training for Ukraine), a flagship instrument of the Atlantic Alliance's direct involvement against Russia.

Since February, Maven has also been deployed in Operations Epic Fury and Roaring Lion, conducted by the United States and Israel respectively against the Iranian regime (and, for Israel, against Hezbollah in Lebanon). It enabled roughly thirteen thousand targets to be struck within the first thirty-eight days of the campaign, at a pace of about one thousand targets in the first twenty-four hours alone — a rate reportedly ten times higher than the best-coordinated U.S. operations of the pre-AI era.[3] With a staff of only twenty operators, Maven carried out a workload that, during the 2003 invasion of Iraq, would have required a hundred times as many personnel.[4]

The compression of the kill chain — the principal objective pursued, and effectively achieved, within U.S. military AI strategy in 2026 — has simultaneously set in motion an evolutionary dynamic and an immense humanitarian challenge: for the great powers, taking seriously the ethical risks posed by curtailing the time available to human operators to assess AI-generated targets, and putting safeguards into practice, would mean imposing on their own defense systems an unacceptable disadvantage vis-à-vis adversaries who disregard such concerns. In the absence of any plausible prospect that international regulation will be implemented while no power has yet secured AI supremacy, the pursuit of ever-greater speed within the OODA loop is consolidating itself as both a selective pressure and a deterrent in its own right.

Intelligentized Warfare

For the Chinese government and People's Liberation Army, the development of military AI is oriented toward the strategic rather than tactical level. Through the concept of intelligentized warfare, China seeks to develop the capacity for its combined forces to wage "systems destruction warfare": not the physical annihilation of enemy troops, but their operational paralysis, achieved by swiftly identifying and striking critical nodes within command, communications, and logistics networks through localized, surgical actions — described through a curious analogy as "military acupuncture." This doctrine also encompasses action within the cognitive domain, on the premise that the human mind and decision-making process are themselves primary battlefields, driving AI competencies aimed at mass disinformation, deepfakes, and personalized influence campaigns designed to corrode the adversary's social cohesion and will to fight. Behind the doctrine of intelligentized warfare lies the ambition to precipitate the collapse of the enemy's defensive and offensive structures even before the kinetic phase begins — though, should armed confrontation prove unavoidable, the doctrine also envisions mass autonomy and drone swarms to saturate enemy defenses.

A significant step toward realizing this doctrine was the sweeping reorganization of the Chinese military structure in April 2024, which dissolved the Strategic Support Force (SSF) and replaced it with three new independent branches, each reporting directly to the Central Military Commission (CMC). The PLA has since comprised four traditional services (Army, Navy, Air Force, Rocket Force) alongside four strategic branches, three created by this reform: the Aerospace Force (space and satellite operations), the Cyberspace Force (defense and offense in digital networks), and — most relevant here — the Information Support Force (ISF). The latter is the fundamental pillar for coordinating network-centric warfare; its digital infrastructure has been described as the "connective tissue" linking the other branches, enabling data flow and fusion across domains (land, sea, air, space, cyberspace). The ISF, through its Integrated Command Platform, seeks to fulfill functions analogous to those provided by Maven — delivering real-time decision support and battlefield data analysis.

Russia: geopolitical constraints and tactical urgency

Russia has made substantial progress toward informational integration across its armed forces, through the Automated Command and Control System (ACCS, or ASU — Автоматизированная Система Управления). Like its Western counterparts, the ACCS was conceived as an end-to-end digital environment, running from detection instruments to command, and from command to weapons, within a single decision-and-execution cycle. Efforts toward its realization began in 2000, with an early breakthrough in 2016–2018: the Unified Tactical-Level Command and Control System (ESU TZ) and its deployment in the Kavkaz military exercises.

With the onset of the "Special Military Operation" in Ukraine in February 2022, the strategic focus on AI use was redirected toward tactical functions — quicker to develop and more immediate in impact. The priority of a unified command-and-control architecture compatible with the American Joint All-Domain Command and Control (JADC2) strategy, for which Palantir is a principal software provider, was set aside in favor of producing and deploying modular software through a bottom-up approach (civilian engineers rapidly developing tools in the field) that solves immediate battlefield problems and accelerates the attack chain in decentralized fashion. The Glaz/Groza systems link drone units to artillery, shortening the strike cycle from hours to minutes and eliminating the need for human radio communication.

The deployment of V2U loitering munitions has also shown significant advances in autonomous swarming tactics. Equipped with NVIDIA Jetson Orin chips, the V2U processes computer-vision and AI algorithms locally, making tactical decisions fully independently. In May 2025, after losing contact with their operator, V2U units reportedly exhibited emergent behavior: the swarm disregarded its planned mission and instead formed a circular holding pattern, from which it developed a plan for simultaneous strikes against vehicles and civilian targets.

Russia, driven by the pressing needs of its Ukraine operations, has distinguished itself through advances in military AI suited to D-DIL environments, where systems must operate under electronically hostile conditions — whether due to the difficulty of guaranteeing a stable communications signal, or saturation of telecommunications bands by enemy EW (Electronic Warfare) operations. In doing so, Russia has to some degree inverted the conventional logic running one-way from strategy to tactics: employing AI modularly at the tactical level, with little connection to advanced control centers, has made it possible to identify high strategic-value targets — command-network nodes, radars, troop concentrations — always in decentralized manner.

Rubber-stamping and ethical tragedies

Given developments of the past three years, the debate over Lethal Autonomous Weapons Systems (LAWS) has shed its status as merely hypothetical to become an urgent ethical problem. Years of deliberation have failed to produce a consensual definition, and attempts within the UN Convention on Certain Conventional Weapons to designate as LAWS those functionally integrated systems capable of identifying, selecting, and engaging targets without human intervention have met fierce resistance from American, Russian, and Chinese delegations. The fact remains that, although fully autonomous systems — with humans entirely absent from the decision loop — do not yet exist, the speed at which information is processed and returned to human agents has come to far exceed the limits of their attention and cognition, creating a situation that, to all intents and purposes, confers de facto autonomy upon the AI system.

Through the Habsora system, employed to identify structures constituting potential legitimate targets in Gaza, and the Lavender system, used to classify individuals by "risk level," the Israeli Defense Forces were able, within the first weeks of the conflict against Hamas and Hezbollah, to identify more than twelve thousand potentially hostile targets; Habsora could flag one hundred structures suspected of sheltering combatants per day, against a historical average of roughly fifty per year.[5] Lavender identified thirty-seven thousand Palestinian men as potential targets.[6] The system relies on Positive Unlabeled Learning, detecting "suspicious" behavioral patterns — such as frequently switching phones or belonging to particular chat groups — applied across the general population. Owing to the overwhelming volume of targets, human analysts have been reduced to mere "rubber stampers," often spending only twenty seconds per target before deciding whether to authorize a lethal strike — wholly insufficient time to reach a decision to employ lethal violence while honoring the principles of proportionality, distinction, necessity, and humanity. At this recognition-and-response speed, with hundreds of strikes possible within a short span, algorithmic warfare leaves an opponent little opportunity to signal de-escalation or negotiate before suffering devastating casualties. The incentive toward preemptive strikes is amplified, since inaction may mean one's own defense systems annihilated within minutes. Fear of being outpaced in decision speed thus encourages preventive force, heightening the risk of accidental or uncontrolled war.

Although human operators are widely believed to be subject to "automation bias" — the tendency to place blind trust in automated systems' suggestions, perceived as "mathematical" and "neutral" even when contradicting clear evidence — the problem appears to stem less from individual psychology than from institutional pressure for performance, exerted to secure decision superiority. Operators face unmistakable hierarchical pressure not to become a source of latency, or be blamed for letting slip a unique opportunity to eliminate an enemy. This can reduce them to perfunctory reviewers — though not willingly. In high-intensity, cognitively saturated scenarios — thousands of targets to discriminate, as many friendly assets to coordinate — human agents, rather than trusting the machine's guidance, tend instead to exhibit "algorithmic aversion": fear of being held responsible for a disastrous act prompted by an AI's judgment.

Thus, the use of military AI in reconnaissance, identification, and targeting appears to occur at the junction of two equally dangerous phenomena: rubber-stamping — where, pressured by results and the inhuman speed of algorithmic warfare, the operator automatically approves the machine's assessment — and operational paralysis, which can arise in contexts of high information saturation and large-scale kinetic attrition, where the operator, driven by algorithmic aversion, introduces a degree of latency that is tactically unacceptable.

What emerges from these three trajectories is not a single race toward the same technological horizon, but three distinct wagers on the same structural imperative: compressing the OODA loop before a rival does. Washington bets on centralized, LLM-driven overmatch; Beijing on the pre-emptive paralysis of enemy systems and minds; Moscow on decentralized solutions forged under fire. What binds these divergent doctrines is a shared absence — any international brake capable of slowing a contest in which ethical restraint is treated, by every major contender, as a luxury only the defeated can afford. It is this race for algorithmic speed, more than any single weapon, that is redrawing the geopolitical chessboards of the multipolar century now under way.

Notes

[1]: Conference on Humanity at the Crossroads. (2024, April 30). Humanity at the crossroads: Autonomous weapons systems and the challenge of regulation [Chair's summary]. United Nations Office for Disarmament Affairs. https://docs-library.unoda.org/General_Assembly_First_Committee_-Seventy-Ninth_session_(2024)/78-241-Humanity-Crossroads-EN_0.pdf

[2]: U.S. Department of War. (2026, January 9). Artificial intelligence strategy for the Department of War: Accelerating America's military AI dominance [Memorandum]. https://media.defense.gov/2026/Jan/12/2003855671/-1/-1/0/artificial-intelligence-strategy-for-the-department-of-war.pdf

[3]: Fitchew, C. (2026, June 8). How AI is transforming conflict and peace. Vision of Humanity. https://www.visionofhumanity.org/how-ai-is-transforming-conflict-and-peace/

[4]: Mande, M., & Allen, G. C. (2026, June 2). What Is Maven Smart System, and What Does It Do? Center for Strategic and International Studies. https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do

[5]: Baggiarini, B. (2023, December 8). Israel's AI can produce 100 bombing targets a day in Gaza. Is this the future of war? The Conversation. https://theconversation.com/israels-ai-can-produce-100-bombing-targets-a-day-in-gaza-is-this-the-future-of-war-219302

[6]: McKernan, B., & Davies, H. (2024, April 3). "The machine did it coldly": Israel used AI to identify 37,000 Hamas targets. The Guardian. https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes