The AI Paradox in Warfare

Decision Superiority or Accelerated Failure?

COMMON EFFORT | AI IN WARFARE | 30 OCTOBER 2026

ABSTRACT

Military AI can increase the speed and scale of intelligence analysis, yet the same integration may weaken the conditions for effective human judgement. This article distinguishes two conditional paradoxes: a control paradox, in which technology intended to improve decisions may encourage dependence and reduce scrutiny; and a strategic-stability paradox, in which faster responses can compress opportunities for verification and de-escalation. Drawing on NATO experimentation and research from CSET, SIPRI, UNIDIR and human-factors scholarship, it argues that military AI should be assessed by decision quality and meaningful human control, not processing speed alone.

Keywords: military AI; paradox of control; decision superiority; command and control; human judgement; strategic stability; NATO

The Race Against Time

In modern warfare, minutes can determine whether a threat is intercepted, an opportunity is lost or lives are saved. Artificial intelligence promises to compress those minutes into seconds. Yet the faster an assessment reaches a commander, the less time may remain to question its assumptions when speed remains the leading criterion. What happens when the information driving a decision is incomplete, manipulated or wrong?

NATO is putting that question to an operational test. At its October 2026 Digital Backbone Experimentation event, personnel from 11 Allies are participating in a Next Generation Targeting experiment involving more than 6,000 simulated targets and over 100 live targets. NATO aims to cut the interval between information and decision by more than half, while retaining ultimate human control over engagement. This is a research objective, not a demonstrated improvement in decision quality.[1]

The distinction is critical. Decision superiority means understanding a situation and choosing an appropriate course of action before an adversary can exploit an advantage; it is not simply processing data first. The paradox is more precise than the observation that faster decisions may be wrong. AI is intended to strengthen command through better situational awareness, yet dependence on its outputs may undermine the time, independent understanding and willingness to question a recommendation that responsible judgement requires. This is the paradox of control: a technology introduced to enhance human decision-making may, under certain conditions, erode the capacity to exercise it.

This tension is not unique to contemporary AI. In her influential 1983 paper Ironies of Automation, Lisanne Bainbridge explained how automation can leave human operators responsible for difficult exceptions while giving them fewer opportunities to maintain the skills and understanding required to intervene. Military AI does not inevitably reproduce this pattern, but the research highlights the difference between retaining formal authority and exercising effective control.[2]

A second, related tension arises at the strategic level: systems designed to increase security through faster recognition and response may reduce the time available for verification and escalation management. Both are conditional, socio-technical paradoxes, not predictions that AI will inevitably diminish safety or decision quality.

Figure 1. The paradox of control: greater AI capability does not automatically secure better human judgement.

The Promise of Decision Advantage

Contemporary operations produce more information than any headquarters can comfortably absorb: satellite images, drone feeds, radar tracks, cyber alerts and intelligence reports compete for attention. AI can filter this material, detect anomalies and help analysts distinguish urgent signals from background noise. A 2025 study by Georgetown University’s Center for Security and Emerging Technology (CSET) identifies genuine advantages for situational awareness, but stresses that benefits depend on clearly defined tasks, training and safeguards against automation errors.[3]

The military value is not limited to targeting. Predictive maintenance can assist with equipment readiness, and data analysis can reveal logistical bottlenecks. In Multi-Domain Operations, information from land, maritime, air, space and cyberspace must also be integrated across organisations that do not necessarily use the same systems. NATO’s 2026 Digital Strategy and the deployment of the Maven Smart System illustrate why interoperability and data quality have become central to command-and-control modernisation.[4][5]

However, a common operational picture is not necessarily a correct one. Integration may remove duplicate reporting, but it can also spread an unnoticed error across several headquarters. The decisive question is whether AI helps personnel understand the picture, rather than merely receive a cleaner version of it.

Figure 2. Operational advantages of AI-enabled decision support.

Ukraine: Operational Gains and Real Limits

Ukraine demonstrates both the potential and the boundaries of military AI. A February 2026 study by the French Institute of International Relations (Ifri), based on open-source material and field interviews, describes AI primarily as an enabler of data processing, target tracking and navigation, rather than an independent battlefield decision-maker. Narrow computer-vision functions can remain useful when electronic warfare disrupts communications.[6]

Reuters documented a Ukrainian drone operation in which AI-assisted guidance reportedly helped a weapon continue towards a suspected Russian tank despite electronic interference. The operator credited the assistance with making the strike possible. Such accounts offer valuable operational evidence, but one successful engagement is not a statistically valid measure of reliability across weather conditions, target types or adversarial countermeasures.[7]

The distinction between AI-assisted guidance and autonomous target selection matters legally and operationally. An algorithm that tracks a target after human authorisation performs a different function from a system that independently chooses whom to attack. Treating both as the same technology obscures the human decisions that remain essential.

The Risk of Accelerated Error

A system may generate a confident recommendation without providing a reliable explanation of how it reached that conclusion. The Alan Turing Institute’s 2025 study of battlefield command warns that AI adoption could reorganise headquarters, information flows and professional judgement; it cannot simply be treated as a faster version of existing staff work.[8]

Imagine an AI tool identifying a concentration of vehicles near critical infrastructure and reporting preparations for an attack. The conclusion combines imagery, intercepted messages and movement patterns. Yet the vehicles might be misidentified, the messages deliberately planted or the movement consistent with several competing explanations. In a compressed decision window, the apparently coherent output may carry more weight than its evidential basis warrants.

Researchers describe automation bias as the tendency to accept automated advice too readily or neglect contradictory information. This is a documented human-factors problem, not proof that military AI systems always cause it.[9] A second problem is common-source dependence: three systems reporting the same erroneous sensor input do not provide three independent confirmations.

An adversary may also attempt to manipulate inputs or exploit model vulnerabilities. The US National Institute of Standards and Technology documents attacks including data poisoning and evasion. In a military setting, these vulnerabilities can become pathways to degraded intelligence, compromised security and unintended operational effects.[10]

Figure 3. AI-related vulnerabilities, including automation bias and data manipulation.

Human Control Is More Than a Signature

Meaningful human control is not established merely by placing a person nominally “in the loop”. A commander may retain final authority while lacking the time, underlying evidence or practical ability to challenge an automated recommendation. When a system handles routine cases and humans intervene mainly in anomalies, they may be asked to make the hardest judgements with limited opportunities to maintain situational understanding. The question is not whether humans must perform every calculation, but whether they retain the knowledge, authority and decision space needed to refuse or redirect AI-supported action.

SIPRI’s research distinguishes autonomous weapons from AI-enabled decision-support systems, while recognising that both can reshape human involvement in targeting. A separate SIPRI assessment explores how biased data and system behaviour can affect the legal requirements of distinction, proportionality and precautions in attack.[11][12] Legal responsibility cannot be delegated to software merely because software helped construct the decision.

There is nevertheless a serious counterargument. Automated responses may offer defensive protection when incoming missiles or drones leave only seconds for action. The policy question is not whether humans should personally perform every computational task, but where human judgement is essential, how the system is constrained and when intervention remains feasible.

UNIDIR has consequently urged policymakers to examine military AI beyond the narrower debate over lethal autonomous weapons, including its wider influence on decisions, international peace and security.[13]

Figure 4. The tension between machine speed and meaningful human judgement.

The Strategic Paradox

At the strategic level, speed may create risks that do not appear in a tactical performance dashboard. SIPRI warns that AI used in conventional intelligence, early warning or targeting can influence nuclear escalation even without being integrated into nuclear weapons. Compressed decision times and opaque recommendations may contribute to miscalculation during crises.[14]

A RAND Europe study commissioned by the UK government similarly explores the wider strategic opportunities and risks of military AI, rather than assuming tactical efficiency translates automatically into stability.[15] One state’s accelerated decision cycle can pressure an opponent to respond equally quickly; the result may be a contest in which neither side has sufficient time to verify intentions.

These are risk scenarios, not evidence that AI has already caused nuclear escalation. They illustrate the second, strategic paradox: a capability intended to improve national security may, if it compresses decision cycles and encourages reciprocal acceleration, undermine strategic stability. Yet faster analysis could also give leaders more time for verification if command procedures protect that opportunity. The effect depends on operational and political choices, not processing speed alone.

Figure 5. Decision acceleration may improve responsiveness while compressing escalation time.

Innovation Versus Restraint

Two competing priorities deserve serious consideration. The operational argument holds that a force unable to exploit AI may lose awareness, responsiveness and resilience against an opponent that does. Delaying reliable innovation could itself cost lives. The precautionary argument stresses that opaque systems, misleading data and weak oversight can turn a technical gain into an operational or humanitarian liability.

NATO’s revised AI strategy provides principles including lawfulness, accountability, reliability, explainability and governability.[16] The Global Commission on Responsible AI in the Military Domain takes the debate further with a responsibility by design approach: controls should be embedded in development, testing, procurement, deployment and institutional practice, not appended only after a system enters service.[17]

Neither position is settled by claims that AI is inherently beneficial or inherently dangerous. The evidence instead suggests evaluating particular systems, in particular roles, under particular conditions.

From Decision Speed to Decision Quality

A defensible assessment of military AI should distinguish decision speed, decision quality and decision control. Accuracy, false positives, confidence calibration, operator workload and the capacity to detect deception matter alongside elapsed time. Tests should include unfamiliar terrain, electronic interference, poor-quality data and situations in which the responsible response is to seek more information rather than act. The capacity to challenge a recommendation is itself an operational performance requirement.

Commanders also need practical AI literacy: not an ability to programme every model, but an understanding of what it was trained to do, where it is likely to fail and what evidence would justify overruling it. A 2026 US Army War College monograph argues that decision advantage depends on the fit between people, models, data and operational context, rather than on processing capacity alone.[18]

Finally, decision processes should protect decision space. Rapid defensive reactions, routine logistics, target validation and decisions with escalation potential do not require identical forms of oversight. A system capable of accelerating all four should not automatically be permitted to compress all four to the same tempo.

Figure 6. Practical requirements for responsible military AI.

The Future: Better Machines or Better Judgement?

Artificial intelligence will remain part of military operations because the operational incentives are substantial. Yet a machine can detect patterns without knowing an adversary’s political intent, and it can provide fluent analysis without possessing the judgement or accountability expected of a commander. Humans are also fallible: they become fatigued, overlook information and succumb to bias. The task is to design cooperation that mitigates the weaknesses of both.

This requires challenging the assumption that acting faster is always acting better. Sometimes speed saves lives. At other times, recognising uncertainty, asking a second question or delaying a decision prevents a far more consequential failure. The ultimate test of military AI is not how quickly it helps a commander decide, but whether it improves the quality and defensibility of that decision.

Figure 7. Decision quality, not speed alone, is the ultimate performance measure.

In the end, the greatest danger may not be that artificial intelligence replaces human decision-makers, but that humans surrender their critical judgement while remaining formally responsible for the consequences. The future of warfare will undoubtedly involve faster machines. Whether it also produces better decisions remains a fundamentally human responsibility.

AI Use and Human Editorial Oversight

This article was developed with the assistance of AI tools for drafting and organization. All editorial decisions and final responsibility for publication remain with the human author. Sources and references should be consulted directly. The accompanying conceptual infographic panels are AI-generated illustrations and are not representations of measured results.

Sources