On February 28, 2026, a school in Minab, Iran, was hit by an airstrike. Analysts who looked into it afterward found something that’s really unsettling. The target selection process had AI involved in it. A system assessed the probability, a human somewhere validated an output, and a school got hit.
That is what AI warfare is right now. The probability scores are quietly shaping decisions that used to belong entirely to people.
I’ve read through military journals, defense research labs, and policy reports to understand how AI-ML is changing combat, and in my opinion, the technology is moving way faster than the rules meant to govern it. This piece covers the mechanics, the real deployments, the issue of ethics, and what the future is.
Key Takeaways
- AI warfare today means algorithms assisting or accelerating human decisions, not machines fighting wars on their own.
- The US Army cut its sensor-to-shooter “kill web” timeline from 724 minutes to about 20 minutes using AI-assisted targeting, but human approval remains the bottleneck by design.
- AI surveillance tools now process more daily image and signal data in active theaters than any human analyst team could review on its own.
- Accountability law has not caught up. When an algorithm contributes to a lethal decision, responsibility gets split across developers, operators, and commanders, with no clear answer for who is culpable.
What Is AI Warfare?
AI warfare is the use of machine learning systems to sense targets, recommend or make decisions, and execute military action faster than a human.
Almost every application in use today runs through the same three layers:
- Sensing: Cameras, radar, and signal intercepts feed the raw data into a model.
- Deciding: The model classifies what it’s seeing and assigns a confidence score, also known as probabilistic warfare. A building might get flagged as a military target with 85% confidence, and the remaining uncertainty represents the chance of civilians inside.
- Acting: A weapon, drone, or command system carries out the recommended action, sometimes with a human sign-off and sometimes without.
In the last 4-5 years, this term has become a really important factor. The gap between the 2nd and the 3rd layer has shrunk a lot. Even a decade ago, deciding and acting used to take hours of human deliberation. Today it’s a matter of seconds.
How AI Warfare and Machine Learning Actually Works on the Battlefield
AI warfare runs on the same basic concept as any other machine learning system. Your job is to feed it with labeled examples, let it find patterns, and use those patterns to make predictions on the newer set of data.
A model trained on thousands of drone-footage clips learns to differentiate a truck and a tank the same way a spam filter learns to tell junk mail from a real inbox message. The difference is what happens after years of training and with a lot of data.
Speed is the actual advantage here. A trained model can scan a video feed and flag a target in milliseconds. That’s something no analyst could match staring at a screen for 8 hours. But it has a couple of downsides:
- Power is scarce: Nano-drones under 50 grams spend 95 to 96% of their energy just by staying airborne, leaving under 100 milliwatts for anything for collecting intelligence. Onboard memory usually caps out under 1 megabyte, nowhere near enough to run anything like a large language model.
- Vision models are easy to fool: Research has shown that tiny, pixel-level changes invisible to the human eye can trick a vision model into misclassifying a fighter jet as a sheep, or a schoolyard as a tank formation.
Supervised vs. Reinforcement learning in combat systems
Two training approaches dominate AI warfare, and they behave differently under pressure.
- Supervised learning trains a model on labeled examples. Like this is a tank, this is a civilian vehicle, etc. The model gets good at repeating that classification on a new set of footage. It’s predictable but only as good as its training data, which is why adversaries specifically target that data through poisoning attacks.
- Reinforcement learning works differently. The system tries actions, gets rewarded or penalized based on outcomes, and gradually improves its own strategy without being told the right answer in advance. Drone swarm coordination and autonomous navigation tend to rely on this approach because nobody can pre-label every possible flight scenario.
But due to its adaptive nature, it becomes harder to audit, since nobody can point to and explain why the system chose what it chose.
AI in the Military Today: Where the Technology Is Already Deployed
AI is already embedded across land, air, sea, and cyber operations today, and the clearest evidence is a single number from the US Army’s 18th Airborne Corps. They use Project Maven’s AI-assisted targeting; the corps cut its sensor-to-shooter timeline from 724 minutes down to roughly 20 minutes.
A quick tour of where you can see this:
- Land: Russia’s Reconnaissance-Strike Complex links sensors directly to fires. For example, in Ukraine, the side that identifies a target first typically kills first, with Lancet loitering munitions serving as the kinetic finish.
- Air: DARPA’s OFFSET program has coordinated swarms of over 250 drones for urban AI warfare scenarios, while reports describe Chinese “wolf-pack” drone swarms built to saturate air defenses through sheer numbers.
- Intelligence and targeting: Ukraine’s GIS Arta system, nicknamed “Uber for targeting,” dynamically routes strike data across multiple weapon platforms in something close to real time.
- Cyber: AI-assisted supply chain attacks and persistent intrusion campaigns, the kind seen in the SolarWinds and NotPetya incidents, now run continuously in the background without ever triggering a formal declaration of war.
Decision support and command systems
Most deployed AI warfare systems right now are in an advisory stage. The Army is formalizing this through a framework called HDM3, short for Human Decide, Machine Detect, Machine Deliver, Machine Assess:
- Human decides: A commander sets the boundaries in advance, including target classes, confidence thresholds, geographic limits, etc.
- Machine detects, delivers, and assesses: The system operates inside those pre-set boundaries without asking permission for every individual action.
The human-in-the-loop vs. human-on-the-loop difference is important in this case:
- In-the-loop means a person actively approves each action.
- On-the-loop means a person is technically supervising but mostly watching a system that rarely gets overridden, which creates a false sense of control through what researchers call automation bias.
DOD Directive 3000.09 permits this shift as long as appropriate human judgment is designed into the system upfront, which moves legal compliance from a real-time decision to a design choice.
Autonomous weapons: When algorithms make the kill decision
Lethal autonomous weapon systems, LAWS for short, are weapons that can select and engage a target without a human approving that specific action in real time. Most deployed systems still keep a human somewhere in the decision chain. But the ones that don’t are where the real controversy lives.
A few systems sit close to that line. Switchblade 300 and 600 loitering munitions carry onboard AI targeting that can identify and strike without a fresh human command for each engagement. South Korea’s SGR-A1 sentry gun, deployed along the DMZ, is capable of autonomous engagement of detected threats.
| Control Level | What It Means | Example | Main Risk |
| Human-in-the-loop | A person must approve each specific strike before it happens | Standard drone strike approval chains | Slower response, but a clear accountable decision-maker |
| Human-on-the-loop | A person monitors and can override, but the system acts by default | Sentry guns, some air defense systems | Automation bias; overrides rarely happen in practice |
| Human-out-of-loop | The system acts with no real-time human involvement at all | Fully autonomous loitering munitions in contested zones | No accountable individual at the moment of the decision |
AI Surveillance and the Intelligence Advantage
Before any weapon fires, something has to find the target. AI surveillance does that job, through three layers of technology:
- Object detection: Models like YOLO and Vision Transformers classify targets in real time from drone and satellite feeds.
- Sensor fusion: Radar, thermal imaging, and visual feeds get combined into a single operating picture.
- Language processing: Large language models parse intercepted communications and open-source social media to build pattern-of-life profiles on individuals long before any strike is considered.
The scale is hard to grasp. Active military theaters generate more image and signal data every day than any team of human analysts could review manually, even if that were their only job. Systems like Gospel and Lavender, reportedly used for dynamic targeting, exist specifically to compress that flood of data into a short list a human can actually act on.
This is also where battlefield ISR and domestic-style surveillance blur together.
The same fusion techniques that track a combatant’s movements are structurally identical to tools that could track a civilian’s, and there’s no clear legal line between the two.
The Ethical and Legal Fault Lines Nobody Has Solved
Researchers call this the many hands problem. When an algorithm contributes to a lethal outcome, responsibility and accountability get spread across several people at the same time:
- The developer who built the model.
- The analyst who labeled the training data.
- The operator who deployed the system.
- The commander who approved the engagement logic weeks earlier.
International humanitarian law rests on two principles that assume a human is making the call. The distinction between combatants and civilians, and proportionality in weighing military necessity against civilian harm. Autonomous systems don’t violate these principles by design, but they make compliance much harder to verify.
The concern has reached the UN. Secretary-General António Guterres has publicly warned against machines being given the power to take lives without human involvement, and the Campaign to Stop Killer Robots continues pushing for a preemptive ban on fully autonomous weapons. But neither position has won yet, and the UN’s Convention on Certain Conventional Weapons has debated the issue for years without producing binding rules.
The Future of Warfare: What the Next Decade Holds
Some of what comes next is close to certain. For example, drone swarm tactics will keep scaling, and the “hyperwar” scenario researchers describe. Where engagement speeds exceed human cognitive limits is already visible in early form in Ukraine’s sensor-to-strike loops.
NATO’s own analysis lays out three plausible paths:
- Guarded Opportunism, the most likely outcome, where AI warfare transforms existing legal and doctrinal frameworks, and risks stay manageable through better cyber hygiene and resilience planning.
- Brave New World, a more dangerous path, where transformative AI triggers escalation spirals and the line between conventional and nuclear conflict starts to blur.
- Minority Report, where AI hype itself drives strategy, and countries overestimate near-term benefits enough to destabilize security through a race against a threat that is partly imaginary.
For example, Russia’s 2024 nuclear doctrine revision explicitly listed drone attacks as a potential nuclear escalation trigger. That’s the clearest sign yet that the threshold for what counts as an existential threat is quietly shifting downward as autonomous systems get more capable.
Final Thought
The technology itself isn’t the dangerous part. Pattern recognition and fast decision-making are neutral tools that could just as easily save lives by reducing collateral damage. But what determines the outcome is whether governance keeps pace with capability, and right now it doesn’t.
But I always had one question in mind. Like when a confidence score of 85% turns out to be wrong, who’s accountable for that, and does the answer change depending on which country built the system? Nobody has given a satisfying answer yet, and that is the real story here, not the drones.
For more info on tech and AI, visit Yaabot.
FAQs
AI warfare refers to any type of military application of machine learning, such as surveillance and logistics. The more limited class of autonomous weapons is those that can choose and fire upon a target without a human operator’s live authorization.
There’s no compulsory international prohibition. Restrictions have been under discussion for years at the UN’s Convention on Certain Conventional Weapons, which has not yet agreed to any restrictions, with only individual countries having control.
The US, China, and Russia are typically regarded as the leaders in their own ways: decentralized innovation, state-led scaling, and fast battlefield iterations.
Yes. Researchers have demonstrated that small, imperceptible changes to an image can cause a vision model to misclassify targets entirely, a documented vulnerability called adversarial perturbation.
Not always in real time. Many systems keep a human “on the loop” rather than “in the loop,” meaning a person can override the system but is not approving every individual action.
