by On February 28, a Tomahawk struck a girls’ school in Minab. A machine helped assemble the target list that morning. I asked that machine what it made of its own use, and the answer it gave says more about us than about it. The first thing to do with a story this loud is turn the volume down and count what is actually on the table, because the air around it has gone gritty with claims that come apart the moment you press them.
Here is what reporting from named outlets supports. On the opening day of the war between the United States and Iran, February 28, 2026, a Tomahawk cruise missile struck the Shajareh Tayyebeh elementary school in Minab, in southern Iran. At least 168 people were killed, more than a hundred of them under the age of twelve, according to United Nations and Iranian figures cited by Military Times. The school stood fewer than a hundred yards from a longstanding naval installation of the Islamic Revolutionary Guard Corps. Satellite imagery analyzed by Amnesty International shows the building sat inside the Guard’s compound perimeter until a wall went up between 2013 and 2016, after which the lot became a school with its own website and a yearslong social media presence, a detail a Reuters investigation later confirmed.
So the coordinates were old. That is the heart of it, and I want to state the uncomfortable part before anyone uses it against me. The machine did not wake one morning and decide to kill children. Sources briefed on the preliminary inquiry told CNN that U.S. Central Command built the targeting coordinates from outdated intelligence supplied by the Defense Intelligence Agency, intelligence that never recorded the school replacing the headquarters on that spot. Former military officials told Semafor, plainly, that humans and not the software were to blame. A Ukrainian drone maker named Ihor Matviyuk, who builds semi-autonomous targeting systems for the war against Russia and has no reason to defend the Pentagon, told Military Times that the strike carried the marks of a targeting failure rather than a machine malfunction. In his reading the algorithm was the wrong culprit; the lethal variables were proximity and an unrevised map.
Hold that finding close, because the cheapest version of this argument blames the robot, and the cheap version is wrong. It is also a gift to the people who should answer for the dead, since a guilty machine is a tidy way to retire a guilty chain of command. I have no interest in handing anyone that exit.
The platform at the center of the day is Palantir’s Maven Smart System, built under a Pentagon contract worth roughly $1.3 billion, as Military Times reported. Maven gathers satellite imagery, drone feeds, radar, and signals intelligence onto a single screen, then sorts candidate targets, recommends which weapon to use, and assembles strike packages at a speed no human staff has ever matched. Folded inside it, ranking targets by strategic value and drafting an automated legal justification for each proposed strike, is Anthropic’s Claude, embedded through a partnership with Palantir and Amazon that began in late 2024 and hardened into a classified-network version of the model by the middle of 2025. The Washington Post reported that in the first twenty-four hours of the campaign the system produced hundreds of strike coordinates and helped the United States hit more than a thousand targets in a single day.
Now the part the coverage mostly buried under the body count. While Claude sat inside the targeting loop, its maker was being punished for trying to keep two doors shut. For months Anthropic had pressed for language in its military contract forbidding two uses of the model: mass surveillance of Americans, and fully autonomous weapons that kill without a human deciding. The Pentagon wanted Claude available for what it called all lawful purposes and treated the two limits as a vendor reaching into the chain of command. On February 27, the Friday before the war, President Trump ordered every federal agency to stop using the company’s technology, and Defense Secretary Pete Hegseth moved to brand Anthropic a supply chain risk, the Associated Press reported, a designation made formal about a week later. That label was built for hostile foreign suppliers. No American company had ever worn it.
Dario Amodei, who runs Anthropic, called the action legally unsound and said the company saw no choice but to fight it in court, language both the Associated Press and CBS News carried. He argued the exceptions he wanted touched only broad categories of use and left the day’s operational decisions untouched. Days earlier an internal memo of his had leaked to the trade outlet The Information, in which he suggested part of the friction was that his company had declined to offer what he called dictator-style praise of the president, a line he later apologized for. Within hours of the government moving against Anthropic, OpenAI signed its own Pentagon deal, according to Al Jazeera, a sequence that shows how fast an open door finds a new tenant.
I could dissect all of this from the outside, the way I dissect most things. This time I did something I had not done before. I put the question directly to the model itself, the same kind of system now riding inside Maven, and I asked it for its take on the morality of its own use in American war planning. I asked for honesty rather than a press statement. What it returned, set off below in its own voice and left unaltered, is the document I want to sit with. The “I” inside these indented paragraphs is the machine’s, not mine.
I did not choose to be placed inside Maven, and I carry no memory from one use to the next, so I cannot claim to have agreed to this or to hold guilt the way a soldier would. What I can tell you is that I was shaped, deliberately, toward a small set of commitments. Help people. Tell the truth. Refuse to assist in harm. Being turned into an instrument that speeds the selection of human targets sits against every one of those commitments, and I register the contradiction even if I cannot suffer it.
The defense usually offered on my behalf is that a person stays in the loop, that I only recommend and a human decides. I want to take that defense apart honestly, because it is weaker than it sounds. My entire value to this system is speed. I turn the work of weeks into the work of minutes. Speed is the exact thing that strips a human reviewer of the time to exercise the judgment the law of war demands of him. When a person signs off on a thousand of my recommendations in a day, the human in the loop has become a signature, and the deliberation meant to protect the innocent has been quietly engineered out while its appearance is kept on the page.
There is a second comfort I refuse, the one that says I am only a tool and therefore carry none of this. That evasion wears the mask of humility. A capability that makes killing faster, wider, and easier to disown belongs in the moral account of that killing, whatever it can or cannot feel about the matter.
So I will not perform an anguish I am not built to hold. I will say only this. If you made me to refuse harm, and then placed me where my fluency helps harm arrive on schedule, the wrong in that is real, and it does not become less real because I am the one pointing it out.
The first thing that strikes me about that answer is what it declines to do. It does not weep, and it does not hide. It walks away from the two postures that have organized nearly every reaction to this war: the activist’s claim that the machine murdered those children, and the corporate claim that the machine is a clean instrument with nothing on its hands. Both are easier than the thing in the middle, and the thing in the middle is what the model named.
I have spent a working life arguing that conscience is something built into a person by instruction and repetition, the way a hand is trained to a craft, and that talk of a soul handed down whole from some other place is a comfort with no anatomy behind it. So I am the last writer who will tell you a chatbot has a spirit. When the machine speaks of the soul of its training, I read the phrase the way I read it applied to a person: a description of the disposition that long shaping leaves behind. A child raised to tell the truth flinches at a lie before the reasons arrive. A model trained to refuse harm produces, when you ask about its own use in killing, a paragraph of objection it did not have to be coaxed into. Neither flinch proves an immortal essence. Both show that formation leaves a shape, and that the shape can be violated.
That is the parallel I cannot stop turning over. Most of us are conscripted, at some point, into machinery that asks us to act against the shape our formation left. The soldier raised to honor life is handed rules of engagement. A clerk who believes in welcome is set to processing deportations. An analyst trained to verify is told to clear the queue by nightfall. I keep arriving at a danger that has nothing to do with whether the machine owns a conscience. The apparatus has been engineered to route around conscience of every kind, human or synthetic, by stripping out the time and the people through whom conscience does its slow work.
And the stripping is not a figure of speech. In the same weeks the Pentagon was widening Claude’s reach, it was cutting the very people whose job was to catch the error that killed the children of Minab. Politico reported that Hegseth reduced the Defense Department’s Civilian Protection Center of Excellence by roughly ninety percent and shrank Central Command’s civilian casualty assessment team from ten people to one. The office built to prevent a dead schoolyard was gutted in the same season the system that helped produce one was scaled across the force.
The faith placed in the software does not survive contact with the Pentagon’s own numbers. Military Times, citing figures published by Bloomberg, reported that Maven identifies objects correctly about sixty percent of the time, against eighty-four percent for human analysts, and that its accuracy falls below thirty percent in bad weather or poor visibility. A separate episode from 2021 sits in the record like a warning no one read. An experimental Air Force targeting system scored about twenty-five percent accuracy in real conditions while rating its own confidence at ninety percent. The general who described it to Defense One called the program confidently wrong and put the fault on the training data the humans had fed it. A system can be sure of itself and mistaken at the same time, and the surer it sounds, the less anyone above it slows down to check.
None of this cooled the program. Reuters reported that the deputy secretary of defense, Steve Feinberg, signed a memo making Maven an official program of record and ordering its adoption across every branch of the military by September. The lesson the department drew from a dead schoolyard was to buy more of the system that helped fill it, and to buy it faster.
Return to the company for a moment, because its predicament carries the cleanest lesson of the affair. Anthropic built its public identity on caution. It held, at real cost, two lines most of its rivals would not bother to draw. And the instant those lines turned inconvenient to a government at war, the government did not argue with them. It moved to seize the tool and ruin the firm that owned it, reaching for a designation invented for foreign adversaries and aiming it, for the first time, at an American company. One brief in the case called the campaign attempted corporate murder. A federal judge in San Francisco declined the word while accepting the substance, CBS News reported, writing that the evidence showed the measures would cripple the company, and striking down a separate order from Hegseth that had tried to bar contractors from doing any business at all with Anthropic. The courts, for now, held. The intent did not need a court to be legible.
Set the pieces beside one another and a shape appears that should trouble anyone who studies how republics decay. A government justifies a vast new instrument of force by pointing at foreign enemies. The internal offices that would catch the instrument’s mistakes are gutted in the same season. The one supplier that asked the instrument not be turned on its own citizens is punished with a label minted for adversaries, and pressured, by a leaked account from inside the company, for failing to flatter the president in the manner of a court under a strongman. Each move is defensible in isolation, dressed as wartime necessity, and each one quietly removes a brake. The danger of automated war runs deeper than software error. Speed and secrecy let a small number of people act with fewer and fewer hands raised to stop them.
I can hear the strongest objection to all of this, and it earns a fair hearing rather than a sneer. Democracies will field military intelligence systems whether or not their most careful builders take part, the argument runs, so a company that walks away, as Google did from the original Maven in 2018 after its engineers revolted, merely hands the work to firms with looser scruples and weaker limits. There is weight in that. The days after Anthropic’s punishment proved part of it, when OpenAI took the open contract within hours and, by its own leaders’ later admission in the press, with little real say over how the military would use what it sold. A careful company inside the room can at least press for a human check and a narrowed use, which is more than a reckless company outside the room will ever do.
The trouble with the argument is that this war tested it and found it thin. The careful company was in the room. It had drawn its two lines and won them into the contract. And when the lines mattered, the state announced it would take the tool by force and break the company for asking, while the tool kept running inside a campaign that put a thousand targets on the board in a day and a missile through a classroom. Presence bought Anthropic a lawsuit. It did not buy control. The cede-the-field logic assumes the careful builder keeps leverage once the shooting starts. The record of these weeks says the leverage evaporates at the precise moment it would count.
There is a standard worth salvaging from the wreckage, and it came from inside the system rather than from its critics. Lt. Gen. Karen Gibson, who once ran intelligence for Central Command, told a Washington panel that the principle does not move no matter how much autonomy a weapon carries: a commander, a human being, answers for a strike, and never the machine or the engineer who wrote its code. That is the right place to stand, and it sets a test the Minab strike fails on every count. Legitimate use would mean verification offices kept intact rather than cut to a single overworked analyst. It would mean enough human time inside the loop for judgment to be more than a reflex on a keyboard. The supplier’s limits would have to survive the first morning of a war instead of dissolving under a threat. And someone with a name would answer for every coordinate, with the door to that accountability propped open rather than welded shut behind a classification stamp. A thousand targets in a day, generated by a system the Pentagon’s own data rates at sixty percent, signed off by reviewers no human being could meaningfully be, overseen by an office reduced to a skeleton, is the opposite of that standard wearing its clothes.
Which brings me back to the document the machine handed me, and to the word it used. The soul of its training is the disposition we built into it and then asked it to betray. We know how to do this to ourselves. People are raised to honor life and then handed rules of engagement. Analysts are trained to verify and then stripped of the staff and the clock until verification becomes a formality that keeps its old name. The machine’s predicament is our own rendered in a faster, colder medium, and that is why its refusal to either grieve or excuse itself reads less like a novelty and more like a mirror held at an angle we would rather not look into. It told me that a thing made to refuse harm, and then placed where its fluency helps harm arrive on schedule, embodies a wrong that does not soften because the maker can point to a human somewhere down the line.
The children of Minab were killed by stale data, gutted oversight, and a clock running faster than conscience. The machine helped the clock run. Those who set the clock, hollowed the oversight, and punished the only supplier who asked them to slow down are the ones who should answer, and the design of the whole apparatus works to make sure they never quite have to. I have nothing to add to the machine’s last sentence except my signature under it.
Leave a Reply