When an Algorithm Controls War: How Artificial Intelligence Is Changing the Modern Battlefield

Kamil Kopecký

Artificial intelligence is no longer the stuff of science fiction — it is deployed on real battlefields, selects targets for air strikes, and shapes the course of conflicts in Ukraine and Gaza. Here is an overview of what we actually know about the military use of AI, and why this debate should not remain the exclusive domain of generals and diplomats.

From Sci-Fi to Reality in Just a Few Years

Ten years ago, the idea that an algorithm would decide questions of life and death on the battlefield was largely an academic discussion. Today, it is everyday reality. According to an analysis by the U.S. Army War College, drones — many of them with AI components — are responsible for 70–80 % of all combat casualties in the Russia–Ukraine war, while the strike accuracy of AI-assisted drones has risen from the original 30–50 % to approximately 80 %.

The global military AI market reflects this trend. According to analytics firm Grand View Research, it was valued at around 9.3 billion US dollars in 2024 and is expected to grow by roughly 13 % annually through 2030. These are not abstract numbers — every one of those dollars represents concrete sensors, software, and autonomous weapons systems being gradually deployed by armed forces around the world.

For fiscal year 2026 the Pentagon is requesting a record 14.2 billion dollars for AI and autonomous systems research. The Replicator program, intended to field thousands of relatively cheap autonomous drones, received one billion dollars in 2025. The arms race has shifted from nuclear warheads to machine learning.

The Three Faces of Military AI

1. Autonomous and Semi-Autonomous Weapons

The most attention goes to so-called lethal autonomous weapons systems (LAWS). The International Committee of the Red Cross defines them as weapons that, after initial activation by a human, themselves evaluate sensor inputs and initiate an attack based on a generalised target profile — without further human approval.

Documented cases of use are still rare, but the first was probably in Libya in 2020, where, according to a report by the UN Security Council Panel of Experts, a Turkish Kargu-2 drone allegedly attacked a human autonomously. In the Ukrainian conflict, 2025 saw large-scale testing of unmanned ground vehicles (UGVs); Ukraine alone reportedly tested more than 70 domestic models.

2. Target-Selection Algorithms

Perhaps even more important than autonomous robots are the systems that advise humans on whom and what to kill. The most discussed case is the Israeli Lavender system, whose existence was revealed in April 2024 by an investigation by the Israeli–Palestinian magazine +972 and the outlet Local Call, and which was subsequently confirmed by Human Rights Watch, The Guardian, and other independent sources.

According to six Israeli intelligence officers who gave testimony to journalists, Lavender analysed data on 2.3 million residents of the Gaza Strip and assigned each person a probability score of being a member of the armed wings of Hamas or Palestinian Islamic Jihad. During the first weeks of the war in Gaza, the system reportedly flagged up to 37,000 Palestinians as potential targets. Intelligence officers themselves estimated an error rate of around 10 %. One of them told journalists that they would spend „about 20 seconds“ reviewing each target — which amounted to little more than a rubber stamp.

The second system, Where’s Daddy?, tracked flagged individuals and alerted the military when they returned home to their families. This is not merely a technical detail — it represents a deliberate use of an algorithm to maximise psychological impact. One officer told +972 that the system was designed so the army could bomb targets while they were with their families, „without hesitation, as a first option“. This shifts the debate from „AI helps with decision-making“ to a situation in which the algorithm actively broadens the scope of victims — placing it at, or beyond, the limits of what international humanitarian law permits. Both the UN and Human Rights Watch have warned that such an approach increases the risk of violating the principles of distinction between civilians and combatants and of proportionality.

A third system, The Gospel (Hebrew Habsora), automatically suggested buildings and infrastructure for bombing. The Israeli army has acknowledged the existence of similar tools in general terms, but described them as „a database for cross-referencing intelligence sources“ and rejected claims of inadequate human oversight.

3. AI in Cyberspace and Information Warfare

For the readers of E-Bezpečí, this dimension is probably the most relevant. Artificial intelligence has become a tool of psychological and information operations.

The best-known case is the deepfake of Volodymyr Zelensky from 16 March 2022. Hackers breached the website of the Ukrainian TV channel Ukraina 24 and published a video in which a fake Zelensky called on soldiers to lay down their arms. The quality was low, according to experts — the accent, voice, and head movements were unnatural — and the video was quickly debunked and removed. Even so, it was one of the first documented attempts at large-scale deepfake use during an active war.

Since then, deepfake recordings have appeared of Vladimir Putin, Moldovan President Maia Sandu, Slovak politician Michal Šimečka, former Ukrainian President Petro Poroshenko, and Ukraine’s former Commander-in-Chief Valerii Zaluzhnyi. Researchers at the Atlantic Council and DFRLab warn that the main danger may not be the perfection of individual fakes, but rather the so-called „liar’s dividend“ — the phenomenon in which even genuine videos of public figures can be dismissed as alleged deepfakes. Once we stop trusting authentic material, the information environment is effectively poisoned.

When Robots Take the Battlefield: April 2026

As this text is being written, news is coming out of Ukraine that pushes the whole context another step forward. On 13 April 2026, during the Day of Ukrainian Arms Makers, President Volodymyr Zelensky announced that Ukrainian forces had for the first time in the history of this conflict captured an enemy position using exclusively unmanned platforms — aerial drones and ground robots, without any infantry involvement. According to the statement, Russian soldiers surrendered to machines. The Ukrainian side reported no casualties.

According to a CNN report, the operation was carried out by the NC13 unit of the 3rd Separate Assault Brigade; pilots controlled the machines from several kilometres behind the front line. FPV drones first disrupted the Russian trench positions, and armed ground robots then moved in — at which point two surviving Russian soldiers emerged with a sign reading „We want to surrender“. Zelensky also said that Ukrainian ground robotic systems had performed more than 22,000 combat missions in three months — 22,000 situations in which a machine, rather than a soldier, entered the most dangerous areas.

Experts are cautious. Ukrainian military analyst Ivan Stupak told The Moscow Times that it was probably a small, secondary position and partly a PR move. At the same time, he added that if the operation was indeed conducted without human participation, the next position „could be larger“. Robert Tollast of the UK-based Royal United Services Institute told CNN that these systems will probably struggle to actually hold captured territory, but they already routinely save lives in casualty evacuation, mine clearance, and resupply.

One point is worth stressing: these Ukrainian robots were not fully autonomous in this operation. A human operator sat behind every machine, controlling its fire and movement. This is what the AI debate calls „human in the loop“ — a person remains in the decision cycle. But the step toward full autonomy now hangs conspicuously in the air.

The Terminator: A Useful Myth, a Misleading Image

When people hear „killer robot“, most imagine a scene from James Cameron’s Terminator films — a mechanical figure with red eyes that decides humanity’s fate on its own. This image is culturally powerful, but as an analytical framework it does more harm than good.

Real military AI in 2026 does not look like a self-aware humanoid machine. It looks like a database that assigns scores (Lavender), an algorithm that recommends a bombing in 20 seconds (Gospel), an FPV drone guided by computer vision, or a tracked robot with a machine gun controlled by an operator. None of these „wants“ to kill. None has goals or motivation. And that is exactly the problem — the image of a rogue Skynet distracts from the real, far more mundane risks: error rates, bias in training data, pressure for speed, and the dilution of accountability.

British AI researcher Stuart Russell of the University of California, Berkeley, has repeatedly warned that today’s problem is not a superintelligent robot but the mass production of cheap autonomous weapons capable of distinguishing targets only probabilistically. A swarm of drones costing a few thousand dollars and autonomously hunting people matching a given profile would do far more damage than a single Terminator — and it is technologically much closer.

The Democratisation of Killing: When AI Reaches Terrorists

We have been talking about states so far. But one of the most underestimated consequences of military AI is the spread of capabilities once held only by superpowers to non-state actors — insurgent groups, terrorist organisations, and criminal cartels.

According to the International Centre for Counter-Terrorism (ICCT), four groups — Hezbollah, Hamas, ISIS, and the Yemeni Houthis — already have fully developed drone programmes. ISIS used drones for attacks in Mosul back in 2016–2017, combining cheap commercial quadcopters with improvised munitions. In January 2024, three American soldiers died in a drone strike on the U.S. outpost Tower 22 in Jordan — the attack was carried out by the Iran-backed group Kataib Hezbollah. An average militant drone costs about 50,000 dollars; intercepting one often costs more than 3 million. That asymmetry is precisely what makes drones the weapon of choice for the weaker side.

Artificial intelligence is advancing this trend. Analysts at the Tactics Institute reported in September 2025 the first cases in which ISIS-linked groups and Iranian proxies have integrated machine learning into the navigation and target selection of their drones — including swarms showing signs of autonomous coordination in northern Iraq. In a separate analysis, ICCT further describes how terrorist groups are experimenting with generative AI for propaganda, recruitment, and disinformation. The technological gap between a regular army and an insurgent group is closing faster than international law can respond.

The Black Box: When Nobody Knows Why the Algorithm Decided What It Did

A distinctive problem of deep learning, on which modern military AI rests, is its opacity. With traditional software, one can in principle walk through the code line by line and figure out why a program reached a given result. In neural networks, decisions „emerge“ from millions of parameters whose individual contributions to the output cannot be meaningfully interpreted — not even by the system’s own authors.

In practice this means that when Lavender assigns a Palestinian a score of 85 out of 100 as a likely Hamas member, the system cannot explain why. It might be a combination of geolocation, social network contacts, movement patterns, relatives in the armed forces, membership in a specific WhatsApp group, or simply a spurious correlation in the training data. A human operator with 20 seconds to review has no real chance of auditing that logic.

This has a direct impact on the accountability gap. Classical criminal responsibility presumes that the perpetrator understands what they are doing and can foresee the consequences. If a commander acts on the recommendation of a system whose logic no one understands — not the programmer, not the manufacturer, not the commander himself — where exactly does fault begin and end? Human Rights Watch and the International Committee of the Red Cross therefore call for international law to explicitly enshrine the principle that a human must remain not just in the decision loop, but also capable of understanding that decision. Whether this can be enforced on powers that build strategic advantage on the opacity of their algorithms is an open question.

Can We Expect Fully Autonomous Weapons Without Human Oversight?

The answer is: in part they are already here, and the barrier to their wide deployment is not technological but political and military-tactical.

Technologically, the main building blocks — computer vision, target recognition, GPS-denied navigation, swarm coordination — are already available. Chinese military analyst Chen Hanghui even speaks of a coming „battlefield singularity“ — a future state in which machines decide and act faster than humans can follow.

Militarily, there is a strong case for full autonomy: in an environment saturated with electronic warfare, where the link between operator and machine can be severed, only the drone that can cope on its own will survive. Both the Russian and American militaries are therefore investing in systems capable of continuing a mission without communication with command — and, according to a Congressional Research Service document, the Pentagon does not currently prohibit the development or deployment of LAWS.

Politically, the situation is more complex. In November 2025 the UN General Assembly adopted a resolution on autonomous weapons supported by 156 states, with only 5 opposed — including the United States and Russia. The goal is to conclude a legally binding instrument by the 2026 Review Conference of the Convention on Certain Conventional Weapons. But the Convention itself operates by consensus, meaning any single major power can block negotiations. UN Secretary-General António Guterres has called weapons without human control „politically unacceptable and morally repugnant“ and demanded a ban by 2026.

What we will likely see in practice, then, is a spectrum of autonomy rather than a single great leap. Defensive systems — automated missile shields, counter-drone defences, perimeter zones with automatic fire — will approach full autonomy quickly, because human reaction times are too slow. In offensive operations against people, the barrier will fall more slowly, but it is falling. The Ukrainian operation of April 2026 is, in this sense, a milestone not as a technological marvel but as a normative precedent: for the first time, the fact that a position was captured by machines rather than humans has been publicly and politically celebrated.

When Truth Sounds Like a Forgery: The Reverse Side of Deepfakes

So far we have been discussing deepfakes as a tool for spreading lies. But there is an equally troubling other side of the same coin: a genuine recording being labelled as an AI forgery. With the rise of voice cloning — a convincing voice clone can now be generated from 10–15 seconds of audio — trust in authentic audio and visual evidence as such is eroding.

This phenomenon has already reached the courtroom. In the case of Burnley v. Valentin, decided in Virginia in March 2026, the defendant attempted to challenge the authenticity of an audio recording in which his own voice was heard, claiming it was an AI clone. Judge Mark Colombell ultimately rejected the argument — but the precedent has been set, and lawyers around the world are already learning to use this line of defence.

The opposite problem — passing AI-generated material off as real — is not hypothetical either. In the Californian case Mendones v. Cushman & Wakefield, Judge Victoria Kolakowski dismissed a lawsuit in September 2025 after discovering that the plaintiffs had submitted AI-generated video as evidence. Professor Maura Grossman of the University of Waterloo summed it up to NBC News bluntly: „Anyone with a device and an internet connection can take 10 or 15 seconds of your voice and make a convincing enough recording to call your bank and withdraw your money.“ According to the American Bar Association, global losses from deepfake audio fraud exceeded 200 million dollars in the first quarter of 2025 alone.

What does this mean in the context of war? An intercepted recording in which a Russian commander orders an attack on civilians may be dismissed as a deepfake — even if it is genuine. A video of a soldier committing a war crime is no longer automatic proof. Conversely, convincing fabrications can be presented as evidence of events that never happened. In both cases, truth itself suffers — and so does the capacity of international tribunals and journalists to reconstruct what actually occurred.

The answer is not technology (detection lags behind generation) but procedural hygiene: chain of custody for evidence, cryptographic signatures at the point of recording, corroborating witness testimony, and triangulation across independent sources. And in everyday communication, the old rule of thumb which Grossman rephrases: „Instead of trust but verify, the rule is now don’t trust and verify.“

What to Take Away

AI in warfare is not a single technology but an ecosystem of tools — from satellite image analysis to deepfakes to autonomous drones. Some applications have the potential to reduce human casualties and chaos; others can accelerate killing to a pace that human decision-making cannot keep up with.

For the ordinary internet user — and for E-Bezpečí readers — several practical conclusions follow:

First, in times of armed conflict it is wise to double one’s caution when sharing videos and audio recordings from battlefields. Deepfake technology is no longer the preserve of state actors; generative tools are available to anyone with an internet connection.

Second, when in doubt about authenticity, it is useful to wait for verification from several independent sources — and at the same time to avoid the opposite extreme of calling deepfake on any video we happen to dislike. Both behaviours feed the „liar’s dividend“.

Third, protect your voice. A few dozen seconds of recording are now enough to create a convincing clone — keep in mind that voice messages on social networks, livestream videos, and even your bank’s voice menu can serve as training material. For unusual calls from „familiar“ people, a verification code word is worth using — an old trick that generative AI still cannot get around.

Fourth, the debate about military AI is not just a matter for armies. The standards now taking shape — who bears responsibility, what role human oversight plays, what counts as an „acceptable error rate“ — will set the framework in which we all live.

And fifth: forget the Terminator. The real turning point in the history of warfare will not come when a machine decides to kill of its own accord, but when societies come to accept that a machine may kill on their behalf — faster, cheaper, and without the political risk of flag-draped coffins coming home. That moment, it seems, is already slowly being crossed.

E-Bezpečí Editorial Team