Every few months, another breathless op-ed asks whether an autonomous drone can be a war criminal. The comment sections fill up with philosophy majors debating robot consciousness, tech ethicists wringing their hands over algorithmic morality, and defense contractors quietly smiling all the way to the bank.
You are asking the wrong question. Worse, you are falling for the oldest trick in the military-industrial playbook. Don't forget to check out our recent post on this related article.
A machine cannot commit a war crime. Not because the code isn't smart enough, and not because we haven't built artificial general intelligence yet. A machine cannot commit a war crime because war crimes require mens rea—criminal intent—and silicon does not have a soul, a conscience, or the legal capacity to stand in the dock at The Hague.
When commentators obsess over whether an algorithm is "guilty," they accidentally build a bulletproof legal shield for the human beings who wrote the code, signed the procurement contracts, and pressed the "execute" button. By anthropomorphizing the weapon, we launder human accountability. We turn a calculated military decision into a "software glitch." If you want more about the history of this, Ars Technica provides an informative summary.
The Lazy Consensus of "Killer Robot" Philosophy
Look at the standard mainstream argument on Lethal Autonomous Weapons Systems (LAWS). It almost always follows a predictable, intellectually lazy script. First, the author defines autonomy. Second, they bring up Isaac Asimov or The Terminator. Third, they argue that because a neural network is a "black box," nobody knows why the drone blew up a school bus, creating an "accountability gap."
This "accountability gap" is a myth. It is a convenient fiction invented by lawyers and defense lobbyists to diffuse blame when civilian casualties inevitably pile up.
Let’s strip away the sci-fi nonsense and look at actual International Humanitarian Law (IHL). Under the Geneva Conventions and Additional Protocol I, the rules of war are crystal clear. Combatants must adhere to three core principles:
- Distinction: You must distinguish between combatants and civilians.
- Proportionality: You cannot launch an attack if the expected incidental civilian harm is excessive compared to the concrete and direct military advantage anticipated.
- Precaution: You must take all feasible steps to spare the civilian population.
Notice who those obligations apply to. They apply to commanders and states. Not rifles. Not artillery shells. Not algorithms.
When an artillery shell lands in a civilian market, nobody convenes a tribunal to ask if the gunpowder had malicious intent. Nobody writes think-pieces about the moral agency of the targeting computer inside a Tomahawk cruise missile. We look at the chain of command. We ask who calculated the blast radius, who authorized the strike, and whether they took adequate precautions to verify the target.
Yet the moment we slap a machine learning model onto a loitering munition, intelligent people suddenly lose their minds. They start treating the software like a rogue mercenary instead of what it actually is: an unguided munition with better sensors.
I Have Seen How Defense Tech Actually Works
I have spent years sitting in windowless briefing rooms watching defense startups pitch "autonomous targeting" to procurement officers. Let me tell you what happens behind closed doors when things go wrong.
Nobody talks about moral philosophy. They talk about liability insulation.
I’ve seen companies blow millions trying to engineer "explainable AI" not because it makes the weapon safer for civilians on the ground, but because it gives the defense contractor a paper trail to hand to government investigators after a strike goes sideways. If an autonomous system misidentifies a water truck as a mobile rocket launcher, the contractor points to the training data, the military commander points to the contractor’s certification, and the politician points to the "fog of war."
Everybody points at the machine. Nobody goes to jail.
This is the exact mechanism legal scholar Madeleine Clare Elish coined as the moral crumple zone. Just as the front end of a car is designed to crumple in a crash to absorb the kinetic energy and protect the human driver, autonomous systems are being positioned to absorb the moral and legal impact of war crimes, protecting the human commanders from prosecution.
If you spend your energy debating whether the machine is a war criminal, you are reinforcing that crumple zone. You are validating the premise that the algorithm is an independent actor.
Dismantling the "Black Box" Excuse
Defense technologists love to hide behind the complexity of deep neural networks. They argue that because modern computer vision models rely on millions of weighted parameters that no single human can trace in real time, a commander cannot reasonably be held responsible for an unexpected targeting error.
That is complete nonsense. In fact, it is an admission of guilt.
Under the legal doctrine of command responsibility—established during the Nuremberg trials and codified in the Rome Statute of the International Criminal Court—a military superior is criminally liable for war crimes committed by subordinates if they knew, or had reason to know, that the forces were about to commit such crimes and failed to prevent them.
Now apply that exact standard to an autonomous weapon.
If a general deploys a neural network into an urban environment knowing full well that the system is a "black box," knowing that its target recognition fails under low-light conditions, and knowing that it cannot reliably distinguish a teenager holding a broomstick from an insurgent holding an RPG, that general is not blameless when the machine massacres non-combatants.
Deploying an unpredictable weapon in a civilian area is not an "algorithmic accident." It is recklessness. Under international criminal law, launching an indiscriminate attack with reckless disregard for civilian life is already a textbook war crime. You do not need to prove the commander wanted civilians to die; you only need to prove they deployed a weapon incapable of making the required legal distinctions.
If I blindfold myself, spin around three times, and fire a machine gun into a crowded street, I cannot blame the gun for not knowing who was a civilian. Deploying an opaque, unverified AI targeting system is the exact same thing, just with a higher R&D budget.
Answering the Questions Nobody Wants to Answer Honestly
When people search for answers about AI warfare, they usually get fed sanitized, academic hedging. Let’s answer the core questions directly, without the diplomatic softening.
Can an AI ever be held legally accountable for a war crime?
No. Legal personhood requires the capacity to hold rights and bear duties, alongside the capacity to be punished. You cannot imprison an algorithm. Deleting a line of code or decommissioning a server is not justice; it is software maintenance. Any framework suggesting we create "corporate personhood" for military AI is a legal trap designed to shield human decision-makers from prison.
Doesn't human-in-the-loop solve the problem?
No, and pretending it does is dangerous. "Human-in-the-loop" sounds comforting to politicians, so they put it in every policy document. In reality, it is mostly theater. When an AI flags a target and gives a human operator three seconds to click "approve" while staring at a degraded video feed in a high-stress combat zone, that human is not exercising meaningful control. They are suffering from automation bias—the well-documented psychological tendency to over-rely on automated systems and assume the computer knows better. The human becomes a rubber stamp whose sole function is to absorb legal blame if the machine makes a mistake.
Are autonomous weapons actually more ethical than human soldiers?
Advocates like roboticist Ronald Arkin have long argued that robots don't get angry, don't seek revenge, don't get tired, and don't commit atrocities out of fear or rage. On paper, that sounds great. In practice, it trades emotional atrocities for statistical ones. A human soldier might hesitate to shoot because they recognize a child's posture, a cultural nuance, or sheer terror on a face. A computer vision model does not feel hesitation; it calculates probability scores. If the threshold is set at 85% confidence and the model hits 85.1%, it pulls the trigger every single time without blinking. Replacing human empathy with rigid statistical thresholds doesn't eliminate war crimes—it industrializes them.
What Needs to Happen Instead
If we want to prevent autonomous weapons from turning modern battlefields into automated slaughterhouses, we need to stop treating this as a philosophical puzzle and start treating it as a strict liability problem.
First, ban the "black box" defense in military procurement. If a defense contractor cannot mathematically verify and explain how a targeting algorithm arrives at a lethal decision, that system is legally defective by default. Under Article 36 of Additional Protocol I, nations are already required to conduct legal reviews of new weapons before fielding them. Right now, most states treat Article 36 reviews as a bureaucratic rubber stamp. If a weapon cannot pass a rigorous, adversarial test for distinction and proportionality across unpredictable environments, its deployment must be classified as an automatic violation of the laws of war.
Second, impose strict criminal liability on the chain of command. We need an explicit international protocol stating that the deployment of an autonomous system transfers zero moral or legal agency to the machine. Every kill executed by an algorithm must be legally treated as if the commanding officer personally pulled the trigger. If a general knows their own freedom is on the line when an algorithm hallucinates a target, watch how quickly the demand for "fully autonomous" lethal systems evaporates. Suddenly, generals will insist on keeping humans firmly in control—not as rubber stamps, but as genuine decision-makers.
Third, hold software engineers and executives accountable for fraudulent reliability claims. In the commercial tech world, "move fast and break things" results in a buggy app or a stock dip. In defense tech, overselling the accuracy of a computer vision model results in dead families. When contractors manipulate test data to hide high false-positive rates in target recognition, that isn't a civil contract dispute. It is aiding and abetting unlawful killings.
The Uncomfortable Truth About Why We Keep Asking the Wrong Question
There is a catch to my argument, and we need to look at it squarely.
If we enforce strict human liability and ban opaque targeting algorithms, Western democracies will move slower than authoritarian adversaries who do not care about the Geneva Conventions, civilian casualties, or war crimes tribunals. Peer competitors will happily deploy cheap, reckless swarms of autonomous drones that shoot anything that moves, gaining a brutal tactical advantage in the first seventy-two hours of a conflict.
That is the real dilemma. Not whether a robot can feel guilt, and not whether an algorithm has a soul.
The real dilemma is whether democratic militaries have the discipline to maintain human accountability when doing so puts them at a speed disadvantage against enemies who have abandoned morality altogether.
Stop asking if machines can be war criminals. Machines don't care about the answer, and the people building them are counting on you to stay distracted by the question. Start looking at the humans hiding behind the code, and decide whether you are going to let them get away with murder.