On 26 September 1983, Lieutenant Colonel Stanislav Petrov was on duty at the command post of the Soviet early-warning system. In the middle of the night the system reported: intercontinental missiles inbound from the United States. First one, then more; the screen showed launches; protocol demanded he report it up the chain — and a report, in that situation, meant a retaliatory nuclear strike. Petrov had seconds. He decided the machine was lying: too few missiles for a real attack, too much like a glitch. He reported a system malfunction. He was right — it was a false alarm, sunlight glinting off clouds that the satellite had read as launches.
One human doubt against the machine's confident testimony — and it may be that this very doubt stands between us and a today that never happened. The entire ethics of war involving artificial intelligence fits into a single question: will there still be even one Petrov in the system — a human capable of disbelieving the machine and saying "no"?
This text is about the algorithm's entry into the most intimate act of war: the choice of whom to kill. About systems that rank and recommend targets. About the difference between a human "in the loop" and a human who has become a signature. And about the accountability gap that opens up when the officer pointing to the target becomes a machine that cannot be hauled before a tribunal.
The whole history of weaponry is the history of widening the distance between the one who kills and the one who is killed. And that distance is not only physical. It is moral.
With a sword you kill feeling the resistance of the body, seeing the face, in the blood. With a spear — one step further. With a bow — at tens of metres, the face already blurred. With a rifle — at hundreds. With artillery — beyond the horizon; you do not even see whom you hit. The drone operator kills from behind a joystick, on shift, and then drives home for dinner. Every rung of this ladder added distance — and every rung made killing psychologically easier, because it grew more remote from human experience.
The algorithm is the next rung — perhaps the last. Here the human no longer even aims. They look at a screen where the machine has already marked the target, assessed it, computed the probability, and they press "confirm." Moral distance becomes nearly total: it wasn't I who found it, not I who decided, I merely agreed with the system's recommendation. The distance regulator is turned all the way up — and it is precisely where almost no resistance is left that the most dangerous zone begins.
Earlier in this series we looked at how the kill chain compressed from days to minutes, and how a surplus of data gave birth to a new kind of fog. Here is the direct sequel: when there is too much data and too little time, an unbearable pressure arises to delegate the hardest link — the choice of target — to the machine.
The logic is relentless. Sensors deliver more potential targets than a human can process. Each one demands analysis: who is this, is it a combatant, is it worth a strike, what are the risks nearby. The human analyst drowns. And then a tempting solution appears: let the machine sift the stream, cross-reference data from different sources, single out those who "fit the criteria," and serve up a ready, ranked list. The human is left only to approve.
Here a special kind of honesty is required, because the subject is poisoned by both propaganda and secrecy. What is known about specific AI systems in targeting comes mostly from journalistic investigations, not from confirmed official descriptions, and every detail remains the object of fierce dispute.
Journalists have reported on Israeli systems allegedly used in Gaza: one for generating a list of object-targets, another for identifying individuals as potential targets on the basis of analysis of large data sets. According to these publications (which the Israeli side has disputed), one system reportedly flagged around 37,000 people, and the human "review" of each target took, by those same reports, on the order of 20 seconds — in effect time enough to confirm the list named a man, not time to make a decision. These figures must be read with all the caveats: they come from disputed journalism, not from confirmed documents.
The honest position is this: we cannot assert precise facts about closed military systems. But even if one sets aside the sharpest accusations, the direction is undeniable: AI is already being used to process intelligence and prepare strike decisions, and the pressure of time and volume pushes the human role from "deciding" toward "approving." It is not the specific system that matters for this text, but the mechanism it makes visible. And that mechanism is universal.
In debates about autonomous weapons, three levels of human control are distinguished. The difference between them is the difference between a conscience and a decoration.
| Position of the human | What the machine does | What the human does | What it becomes under pressure |
|---|---|---|---|
| In the loop | proposes a target | must explicitly decide — without them there is no strike | a stamp: automation bias turns the decision into approval |
| On the loop | acts on its own | supervises, can intervene | rarely manages to intervene in time |
| Out of the loop | detects, decides, strikes on its own | nothing | full autonomy of the lethal decision |
The central ethical and diplomatic battle is fought around this — over "meaningful human control." Everyone publicly swears that the human will remain in the loop. But the devil is in the word "meaningful." Because there is a way to keep the human in the loop formally while turning them, in fact, into nothing. And that way has a name.
The name is automation bias. It is a well-documented psychological effect: when a machine gives a confident recommendation, people tend to obey it — even against their own doubts, even when the machine is wrong. Studies in aviation and medicine record that operators, again and again, follow the system's mistaken prompt, ignoring the correct readings right in front of their eyes. The more complex the system and the greater the pressure, the more the human "switches off" their own judgement. In essence, this is the Milgram experiment without the man in the lab coat: the role of the authority to be obeyed has been handed to a machine wreathed in an aura of precision — and it is obeyed all the more readily.
Now overlay this onto war. An analyst before whom the machine has placed a hundred targets tagged "high probability," exhausted, time-pressured, surrounded by trust in an expensive, intelligent system. How many seconds will they spend on each? (Recall the disputed "20 seconds.") Are they really checking — or pressing "confirm," because the machine rarely errs, because there's a queue, because everyone does it this way, because who are they to argue with a billion-dollar system? Formally, they are a human in the loop. In fact, they are a stamp that legitimizes the machine's decision without changing it.
This is why the mere fact that "a human approves" is not enough. If they have neither the time, nor the data, nor the psychological footing to genuinely disagree — their presence in the loop is a fiction, a moral fig leaf. They are needed not to control the machine, but so that there is someone on whom to lay responsibility and someone to shield a decision that was, in truth, made by the algorithm.
And here we arrive at the deepest problem, one Hannah Arendt described long before any AI. Watching the Eichmann trial, she introduced the notion of the "banality of evil": the most monstrous things are done not by monsters, but by ordinary people embedded in a bureaucracy where each performs their narrow step and no one feels responsible for the whole. Responsibility dissolves into procedure.
Algorithmic killing is the banality of evil refined to engineering perfection. Who answers for a mistaken death? The operator: I trusted the system, it had high accuracy. The commander: I approved the procedure, not the specific target. The developer: I wrote a tool, the decisions were made by humans. The machine will say nothing — it cannot be interrogated, convicted, punished. Responsibility is smeared along the chain so thinly that no stain is left on anyone. An accountability gap forms: there is a death, but structurally there is no one guilty.
This is worse than Arendt's classic bureaucracy, because there you could still point a finger at the human who "was only following orders." Here the order was given by a statistical model, and the people around it sincerely believe they had nothing to do with it — each performed a small technical step. Evil became not banal, but distributed — and therefore almost impossible to grasp.
There is one more pressure against Petrov — the very one Virilio wrote about and the one discussed in the text on the kill chain. The faster the strike cycle, the less room there is in it for human judgement.
When detection to strike is minutes or seconds, the human physically cannot manage to make sense of it. And an adversary who has removed the human from the loop acts faster still — and wins the tempo. A race of autonomization arises: whoever leaves more decisions to the machine is faster, and whoever is faster wins. Each step toward autonomy is rational on its own and catastrophic in sum — the classic trap. It is exactly this "doomsday" machine, acting on logic without a human, that terrified audiences in Dr. Strangelove and WarGames: a computer that cannot tell war from a game until it deduces that "the only winning move is not to play." We are building a war in which no time at all is allotted for Petrov's doubt.
It would be dishonest to paint everything one-sidedly. There is a serious argument in the machine's favour, and it is worth stating in its strongest form.
The human at war is a poor judge. They are tired, frightened, vengeful. They shoot in panic, err from exhaustion, kill out of rage over a fallen comrade. The machine does not fear, does not avenge, does not tire at four in the morning. In theory, a well-tuned system can be more accurate and more restrained than a human — fewer accidental casualties, fewer emotional decisions, stricter adherence to criteria. Its proponents say: if the machine errs less often than a frightened soldier, then refusing to use it will cost more innocent lives, not fewer.
The argument is strong — and still it does not close the question. First, the machine's accuracy depends on the data it was trained on, and that data carries all the biases of those who gathered it; an "objective" machine can systematically err in one direction — confidently and at scale. Second, even a perfectly accurate machine does not solve the problem of accountability: who answers for its mistake? Third, and chief: the question is not only the number of casualties, but what we lose when we remove human doubt from the decision over death. Petrov was "inefficient" from the standpoint of protocol. It may be that his very inefficiency saved the world. The machine is never inefficient for humanity's sake against its instructions — it does exactly what is wired into it.
The delegation of decisions to the machine left the bounds of war long ago — into medicine, credit, hiring, justice. The questions worth asking yourself wherever an algorithm "recommends":
The algorithm enters war not as a fantastical terminator, but mundanely — as a handy assistant that sifts the data, ranks the threats, compiles the lists, and saves precious minutes. Each step of this delegation is rational. The sum of the steps leads to a world where the machine picks the target and the human merely signs, so that there is someone to assign the responsibility that, in truth, no one bears.
The danger is not that the machine will "rise up." It is more prosaic and closer at hand: that we ourselves, under the pressure of speed, volume, and competition, will voluntarily reduce the human to a stamp — and call it "human control." That automation bias will quietly turn every analyst into a stamp that sincerely believes it is still deciding. That the accountability gap will swallow guilt so smoothly that no one will even notice when, for the last time, someone genuinely decided.
In 1983 the world was saved not by a better machine, but by a worse subordinate — a human who broke protocol and refused to believe the screen. The entire ethics of war with artificial intelligence comes down to one question, and it is not a technical one: have we left in the system a place for Petrov — for the one who has the time, the right, and the courage to tell the machine "no"? Because an army that has fired all its Petrovs for inefficiency will one day press "confirm" on a false alarm — and there will be no one left to doubt.