AI in Warfare: Real Uses, Risks and the Human Cost

When War Learns to Think

Artificial intelligence is entering conflict through intelligence, defense, logistics, information and decisions—not only through weapons. Here is what it really changes, and why human judgment matters more than ever.

Artificial intelligence is transforming warfare long before a weapon is fired—changing how military information is analyzed, understood, and turned into decisions.

An algorithm designed to study endangered birds is an unlikely place to begin a story about modern war. Yet in June 2026, NATO’s Allied Command Transformation described how research originally used to analyze bird migration patterns was being adapted to identify patterns in Russian military movements. The model was being integrated into Ukraine’s DELTA battlefield information ecosystem. The platform was already widely used; the specific research model had not yet completed that integration. That distinction matters. NATO’s account of the project is a genuine example of military AI, but not a claim that a scientific prototype was already directing combat.

The story reveals something easily lost beneath dramatic headlines about autonomous drones. The same mathematical techniques can track animal populations, predict traffic or help analysts make sense of the movements of an army. The algorithm does not know whether it is studying birds or soldiers. People decide what information to feed it, what patterns to search for and, most importantly, what to do with its conclusions.

This is where the relationship between artificial intelligence and warfare begins. Before AI can change how a weapon operates, it can change what a military sees, how fast information circulates and what a commander believes is happening. It can help protect a city from an incoming attack, move supplies to an exhausted unit or amplify a false video before anyone has checked whether it is real.

For most of us, AI means a chatbot, a photo editor or a recommendation system. At war, the underlying technology enters an environment where information is incomplete, people deliberately deceive one another and mistakes cannot simply be undone. To understand AI in warfare, we should first look past the machines that fire and examine the much larger system that decides, organizes, protects and persuades.

Why does war need artificial intelligence at all?

Modern warfare is not a video game with a neat map showing every friendly and hostile unit. It is a struggle to understand a changing reality. Drones transmit video. Satellites collect imagery. Radar stations produce detections. Troops send reports. Communications fail, sensors contradict one another and information grows old while someone is still reading it. A military might know more individual facts than at any point in history and still fail to see the larger picture.

AI is attractive because it promises to do some of the sorting. Machine-learning software can recognize recurring visual patterns, help correlate reports from different sources, flag unusual behavior and identify where a human analyst should look next. The value is not magical foresight. It is the possibility of turning an overwhelming stream of disconnected signals into a more understandable, more timely picture.

A simple civilian analogy helps. A navigation app does not drive every car or control every traffic light. It examines a vast amount of information and suggests a route. That suggestion can be useful, but it can also be outdated, based on a wrong assumption or unsuitable for the person taking it. Military decision-support AI operates under much harsher conditions, with an intelligent adversary trying to contaminate the information and with far greater consequences for being wrong.

It is also important to separate three terms often thrown together. Ordinary automation follows a defined set of instructions. Machine learning extracts patterns from data and applies them to new examples. Generative AI produces text, images, code or other content based on what it has learned. These approaches overlap in real systems, but not every automated drone uses modern AI and not every use of AI has anything to do with a drone. Our plain-English introduction to AI and neural networks explains the underlying technology in more detail.

In short, military AI is not one invention. It is a collection of tools entering an organization already built around intelligence, uncertainty, logistics and command. The tools matter because they can influence a decision long before the visible moment of action.

The invisible battlefield: where AI is already useful

Seeing the battlefield without watching every frame

Imagine a reconnaissance team receiving many hours of footage from different cameras. Human analysts can study it carefully, but they cannot watch every recording in real time, remember every object and compare all the frames from previous days. Computer vision can help mark potential vehicles, changes in terrain or repeated movement, giving analysts a list of leads rather than an unmanageable archive.

The US Army describes its Maven Smart System as a way to bring together information, including imagery and video, to improve situational awareness and support commanders. Ukraine’s DELTA ecosystem performs a different but related role as a battlefield information and coordination environment, with AI-enabled capabilities being added around it. Neither example should be reduced to the sensational claim that a single machine is independently running a war.

Yet the limits are just as important as the possibilities. A flagged shape is not proof of a hostile vehicle. A location may have changed since the image was recorded. A civilian object may resemble a military one. Good analysts do not merely receive a prediction; they ask how the system arrived at it and what else could explain the evidence.

Making sense of a situation, not replacing a commander

Once information has been collected, the next challenge is interpretation. AI tools can help organize reports, compare possible scenarios, identify resource shortages or show which assumptions a plan depends on. A language model might summarize a long technical document or translate information more quickly. A specialized model might search historical patterns to identify a change worth investigating. This is decision support, not independent military judgment.

There is a subtle danger in even the phrase “decision support.” The form in which software presents a situation can shape what the human thinks is important. If a map highlights three locations and hides uncertainty about a fourth, the officer may spend valuable time on the highlighted options. Nobody needs to surrender formal command to a machine for a machine to influence the decision.

Protecting people and critical infrastructure

The defensive use of AI can be less spectacular and more immediately understandable. A city facing waves of drones must distinguish real threats from irrelevant signals, combine alerts from different sensors and distribute limited defensive attention. Pattern-recognition systems may help with parts of that work. They do not guarantee successful interception, and effective defense still depends on many non-AI factors, including trained personnel, reliable communications and enough equipment.

The same logic extends beyond the sky. Cybersecurity systems can help security teams notice unusual activity in large networks, while analysts investigate whether it is a real intrusion or a false alarm. That matters when electricity, healthcare and public communications are under attack. In a war affecting civilians every day, earlier warning or clearer coordination is not merely an abstract improvement in military efficiency. It can be a way to reduce harm.

Logistics, maintenance and medical support

Wars are sustained by surprisingly ordinary questions: where are the supplies, which vehicles require repair, what equipment is likely to fail and how can a dangerous journey be avoided? Statistical forecasting and machine learning can assist with maintenance schedules, inventories and transportation planning. These uses receive fewer headlines than autonomous weapons, although a functioning supply system can be decisive for the people relying on it.

Remote and partially autonomous ground systems are also being explored for resupply and casualty evacuation, where conditions make the trip particularly hazardous. They are not a substitute for medical teams or a promise that robots can safely handle every battlefield emergency. But the motivation is recognizable from civilian applications such as medical delivery drones: can technology complete one dangerous or time-sensitive task without exposing another person to unnecessary risk?

Here a larger truth comes into view. AI may influence a conflict without identifying a target, choosing a weapon or even being visible to soldiers. A better inventory forecast, a warning delivered earlier or a recovered communications system can change what happens on the ground.

Modern warfare generates enormous amounts of information. AI can help military analysts identify patterns and prioritize data, but understanding the battlefield still requires human judgment.

Ukraine shows why military AI is not just a laboratory experiment

Russia’s full-scale invasion has made Ukraine a particularly revealing case. The pressure is immediate: repeated attacks on cities and infrastructure, intense surveillance, electronic interference, rapid technological adaptation and the need to keep people alive with finite resources. AI is arriving here not as a fashionable addition to a peacetime technology plan, but as one set of tools among many that may help defenders react more effectively.

In August 2025, Ukraine’s government announced that DELTA had been introduced across the Defense Forces. The important fact is not that DELTA should be mistaken for a single artificial intelligence model—it should not—but that a common digital environment allows observations and operational information to be shared and used. The Avengers computer-vision system offers a more direct example of AI: automated recognition of military equipment in visual material. In January 2026, the Ministry of Defence also announced Brave1 Dataroom, a protected environment for developing and evaluating defense models using relevant data.

None of those announcements proves that every model works flawlessly in combat. Public descriptions often come from institutions that develop or deploy the systems, and independent performance evidence can be limited. Still, together they illustrate a practical lesson: the useful AI project is often less about a revolutionary algorithm than about collecting good data, protecting that data, connecting it to the work people actually do and updating the system when the conditions change.

The NATO bird-migration project is an especially revealing example. Civilian research was not invented for war, yet its methods could be adapted for military analysis. That does not make ordinary science morally identical to combat. It shows that the boundary between civilian and defense technology is increasingly porous. Universities, software firms and researchers can contribute capabilities whose eventual use was not part of the original scientific question.

Ukraine also exposes the limits of technological optimism. Software cannot hold territory, replace political decisions or manufacture experienced personnel out of nothing. It can make a particular task easier, but it exists inside a complex human and material system. A country under attack has powerful reasons to pursue defensive innovation; it also has powerful reasons to demand honesty about what that innovation can and cannot do.

What about AI-powered weapons? A short introduction

Weapons deserve a place in this discussion because they are the most consequential boundary of military AI—but they are not the whole story. Computer vision can help a drone follow an object. Some systems can navigate or perform a limited function when communications are disrupted. Defensive interceptors may automate parts of tracking and movement. These are different levels of autonomy, and they should not all be described as machines independently deciding whom to kill.

Reuters reported in November 2025 on Ukrainian interest in AI-assisted drone guidance, while distinguishing autonomous functions from human authorization of attacks. On the Russian side, a 2026 CSIS analysis documented efforts to connect drones, AI research, manufacturing and battlefield feedback. Both examples make a broader point: autonomy is developing in specific tasks, sometimes with inexpensive components, and its real-world effectiveness cannot be judged solely from demonstrations or promotional claims.

Popular discussion also tends to treat every large formation of drones as an intelligent “swarm.” But sending multiple drones into an operation is not the same as a group of machines independently cooperating through sophisticated collective intelligence. RUSI’s 2026 discussion of swarm ambitions helps explain why prototype concepts, supervised multi-drone operation and genuinely collective autonomy must be kept separate.

A later Next Horizon article can examine the technologies in detail—AI-guided drones, autonomous interception, robotic vehicles, sensors and the practical differences between remote control and autonomous operation. For this introduction, the essential question is simpler: which tasks are performed by software, which decisions remain with people and how can anyone verify that the boundary works as intended?

Another front: information can become a weapon

The battlefield does not stop where troops are deployed. Governments, civilians and international audiences see videos, photographs, messages and claims about a conflict. Generative AI can make convincing fake voices and images cheaper to produce, and it can flood online spaces with material that is difficult to check at speed. Not all wartime disinformation uses AI, of course; the underlying techniques of propaganda and manipulation are much older.

In a 2025 analysis, Ukraine’s Center for Countering Disinformation documented AI-generated videos used in efforts to discredit the country’s Defense Forces. Separately, RAND’s 2025 research discussed how generative AI might be acquired for US military influence activities, emphasizing the organizational issues surrounding its use. The cases are different in purpose and status, but together they show why language and image models belong in any serious discussion of defense technology.

This introduces a peculiar modern problem: people may witness more images of war while becoming less certain what they have actually seen. AI can assist fact-checkers and investigators, but it can also make fabricated evidence faster and cheaper to create. Trust itself becomes contested terrain. That is not an argument that nothing can be believed. It is an argument for evidence, provenance, independent reporting and a habit of asking where an image or claim came from.

For civilians, the importance of military AI extends beyond the battlefield. Better threat detection, early-warning systems, and reliable information could help protect lives during wartime.

The global race: armies, governments and technology companies

It would be a mistake to treat military AI as an unusual experiment confined to one battlefield. The United States is incorporating AI into information systems and command workflows. NATO has published principles for responsible military AI, including lawfulness, accountability, explainability, reliability, governability and bias mitigation. China has advanced the concept of “intelligentized warfare,” envisioning closer integration of sensing, software and military decision-making. These ambitions are real policy directions, though a stated vision is not the same as proven operational capability.

Commercial technology is part of the story, too. The US Department of Defense announced a 2025 prototype award connected with OpenAI and another involving Anthropic. These agreements reveal how models and expertise developed for wider markets are being drawn into defense work; they are not evidence that a commercial chatbot has been placed in autonomous control of a weapon. Civilian AI companies face questions about access restrictions, acceptable customers, independent oversight and the social consequences of dual-use technologies.

The strategic competition is not simply about which nation buys the biggest AI model. It is about electricity, chips, skilled personnel, secure networks, trusted datasets and institutions capable of learning from mistakes. The militaries that integrate technology well may gain more than those that collect the most impressive prototypes. This is as much an organizational transformation as an engineering one.

Can we trust AI when the consequences are irreversible?

There is a tempting way to describe an algorithm: unlike a tired or frightened person, it does not panic, become angry or lose concentration. In narrow tasks, automation can indeed reduce some human errors. But it can introduce different errors, sometimes with an especially dangerous appearance of objectivity. A model trained on clear images may struggle with smoke or damaged objects. A report generated by a language model may read beautifully while omitting a vital caveat. An opponent may deliberately try to fool a recognition system or poison its inputs.

The question should never be “Is the AI accurate?” in the abstract. Accurate compared with what, under which conditions, and with which consequences for different kinds of mistakes? A system might be excellent at finding patterns but poor at distinguishing their significance. It might provide an impressive average score while failing on rare conditions that matter most. That is why military claims require more than a demonstration video or an isolated performance figure.

The human-machine relationship is also more complicated than a familiar warning about people blindly following computers. A July 2026 study in the Journal of Conflict Resolution tested a reconstructed military AI decision-support interface in two experiments involving 2,015 Israeli military personnel. The researchers found notable skepticism toward AI-attributed recommendations, particularly when civilian harm was high; explanation features could change how participants evaluated the advice. These were experimental decisions, not observations of actual combat orders, and the findings should not be generalized to every military.

That distinction is valuable. Some people may overtrust AI, others may distrust it even when it is helpful, and both responses can be dangerous. Merely putting a human name on an approval screen does not guarantee wise judgment. The person needs access to the evidence, an honest indication of uncertainty and the real ability to disagree. The better model of military AI is not a machine replacing human thought, but a system whose design makes human thought more informed without making it ceremonial.

When an AI suggestion becomes a question of life and death

The debate is no longer hypothetical. An investigation published by +972 Magazine and Local Call in April 2024 described allegations about an Israeli military system called Lavender and its role in generating targeting recommendations during the war in Gaza. The Israeli military disputed key claims, describing the relevant system as a tool supporting analysts rather than an autonomous system selecting targets. The competing accounts should not be flattened into a settled technical description of exactly how every decision was made.

What the controversy does illustrate is a broader dilemma that applies well beyond one conflict. A list of recommendations can affect operations even when a person formally retains the final decision. If software produces suggestions faster than humans can properly investigate them, the problem is not only how the algorithm works. It is also how an institution defines acceptable evidence, how much time it allows for review and whether rejection is genuinely possible.

I find this more troubling than the image of a fictional robot soldier. A machine that visibly acts on its own would at least make the loss of human control obvious. The harder situation is one in which the institution can truthfully say that a person clicked “approve,” while no one can explain whether that person had sufficient knowledge or freedom to make an independent decision.

AI can recognize patterns and recommend actions, but it cannot replace human responsibility. As military decisions become increasingly automated, understanding uncertainty and accountability becomes essential.

Could smarter military systems make wars more dangerous?

Speed is usually presented as a benefit. Detect a threat earlier, respond faster, complete a task with fewer delays. In defense, those seconds may save people. But strategy is not always improved by moving faster. A system can send an urgent warning before the underlying evidence is settled; another country may read a defensive response as preparation for an attack; politicians may feel pressure to choose before diplomats and intelligence analysts have time to examine the uncertainty.

The Stockholm International Peace Research Institute warned in 2025 that military AI can influence nuclear escalation even when it is used outside nuclear weapons themselves. The risks involve compressed decision times, changing threat perceptions and possible effects on strategic stability. Research with language models has also explored escalation in simulated crises, including a 2024 preprint and a 2026 preprint. These are models of hypothetical decision settings, not reliable predictions of how actual governments will behave or proof that AI has independently brought states close to nuclear war.

The deeper danger is a familiar human temptation: to confuse additional information with certainty and rapid calculation with wisdom. A computer can optimize for a specified objective. It cannot determine by mathematics alone whether that objective is politically justified, legally permitted or morally worth its cost. The more capable the tool, the easier it becomes to forget who chose the objective in the first place.

Does international law still apply to AI in war?

Yes. New software does not suspend the laws of armed conflict. International humanitarian law requires parties to distinguish civilians and civilian objects from lawful military targets, to assess proportionality and to take feasible precautions in attacks. Those obligations fall on humans and states, not on software treated as if it were a legal or moral person. A highly precise technology can still be used unlawfully; accurately striking the wrong target is not a success.

The International Committee of the Red Cross defines autonomous weapon systems in terms of selecting and engaging targets after activation without further human intervention and warns that existing law does not answer every new practical or ethical question. In August 2026, the UN and ICRC renewed calls for additional binding rules. No universal treaty has solved the entire issue. The legal debate is consequently about both enforcing existing obligations and defining limits for systems that make specific lethal decisions without further human input.

A sound principle should be concrete rather than ceremonial. If a human is said to control a system, can that person understand the relevant situation? Can they challenge the output or stop the action? Have realistic failures been tested, including the possibility of deliberate deception? Can investigators reconstruct the decision afterward? The US Department of Defense’s autonomy policy and NATO’s responsible-use framework address elements of this challenge. Policies matter, but their real test is implementation.

The question becomes especially painful for a society defending itself against aggression. Refusing potentially protective technology can cost lives today. Adopting technology without credible limits can impose other costs tomorrow. Those pressures do not cancel each other out. They are the reason that responsible development has to be practical, not merely an abstract declaration that humans should stay in charge.

What the next five years could look like

The most plausible near-term change is not a battlefield filled with science-fiction robot armies. It is AI becoming a less visible but more ordinary part of military activity. Analysts may rely on more automated image triage. Command staffs may receive faster but more complicated recommendations. Defense systems may combine signals from more sensors. Logistics and maintenance teams may use predictive software while trying to keep essential operations running through disrupted networks.

Those developments will also create new dependencies. A force may depend on commercial chips, foreign cloud services, proprietary software updates or data it cannot easily audit. An algorithm that once helped could become a liability if an adversary discovers how to mislead it. The less visible AI becomes, the more important it will be to know where it sits in a decision chain and what happens when it fails.

There are encouraging possibilities. Better warnings could reduce civilian exposure; reliable remote systems could keep people away from unnecessary danger; better analysis could prevent some mistakes and better scrutiny could strengthen accountability. There are darker possibilities as well: cheaper mass surveillance, more persuasive wartime fabrications, faster strikes and institutions hiding responsibility behind machine-generated recommendations. Technology will not select one of these futures on its own. Public oversight, law, organizational culture and political choices will help decide which direction becomes normal.

Conclusion: what should machines be allowed to decide?

I am fascinated by the idea that an algorithm designed to understand migrating birds can someday help analysts understand the movements of an army. It demonstrates the extraordinary reach of modern scientific tools. It also forces an uncomfortable realization: a technology does not carry a moral purpose inside its code. Its purpose emerges from what people ask it to do, how they test it and what they allow to happen on the basis of its answers.

Artificial intelligence will not eliminate the human causes of war. It cannot resolve territorial ambition, erase aggression or settle questions of justice. But it can already alter the conditions in which people act: what they know, what they think they know, how quickly they move and how confidently they justify what they have done. Some of those changes may save lives. Others may make violence easier to scale and harder to question.

Perhaps we are asking the wrong question when we wonder whether AI will someday be intelligent enough to fight a war by itself. The immediate challenge is more subtle. As systems become better at finding patterns and suggesting actions, will the people using them become better at understanding consequences—or simply faster at accepting answers?

In almost every other field, we applaud AI when it can do more without human help. War should make us think carefully about that instinct. The measure of progress may not be how many decisions a machine can make, but how reliably it helps human beings make the decisions they must never abandon.

The future of military AI will depend not only on what technology can achieve, but also on what humanity decides it should be allowed to do.

FAQ: AI in warfare and defense

How is AI used in warfare today?

AI is used or tested for analyzing images and sensor data, organizing intelligence, supporting command decisions, detecting cyber anomalies, planning logistics and enabling limited autonomous functions. The maturity and evidence behind individual systems vary widely.

Is military AI the same as autonomous weapons?

No. An AI system might analyze a satellite image or organize supplies without controlling any weapon. Autonomous weapons represent a narrower and more legally sensitive category because a system may select and engage targets after activation without further human intervention.

How is Ukraine using artificial intelligence?

Publicly documented examples include AI-enabled computer vision, the DELTA information ecosystem and protected efforts to train and test models. Reports also describe limited autonomous functions in drones. Military statements and battlefield performance claims need careful attribution.

Can AI help protect civilians during war?

It may support threat detection, early warning, cybersecurity, maintenance and emergency coordination. Its effectiveness depends on the larger system and its reliability; AI does not, by itself, make an operation safe or lawful.

Can AI make mistakes in military decisions?

Yes. Models can misread unusual conditions, rely on outdated data, miss context or generate convincing but false summaries. Human review must be meaningful, informed and able to challenge recommendations.

Will AI replace soldiers and military commanders?

There is no sound basis to expect complete replacement in the near term. More tasks are likely to be automated, but strategy, legal accountability, human relationships and the complex physical realities of conflict do not disappear.