Outdated: The Word That Killed 175 Children

On a morning in early 2026, a primary school in the Iranian town of Minab was struck from the air. At least a hundred and seventy-five people were killed, most of them girls between seven and twelve years old. The United States' own preliminary investigation, reported by the New York Times, found the strike was American and attributed it to outdated targeting data. Not a wrong decision. Not a misidentification. Old information, in a system built to move faster than anyone could check it.

By A. Kade

I want to start with the word, because the word is the whole piece.

Outdated.

It is a word you use about a phone, or a licence, or a list of contacts you have been meaning to clean up. It is a mild word. It carries no blame in it at all. Nobody is outdated. Data is outdated, and data does not have a father, or a Tuesday, or a school bag on a hook.

A hundred and seventy-five people died at Sharejeh Tayebeh Primary School in Minab. Around a hundred more children and staff were wounded. Red Crescent medics and parents described a second strike following the first. Investigations by the BBC, NPR, CBC and the New York Times all concluded American forces most likely carried it out, and the United States' own preliminary inquiry agreed.

And the reason given, the technical, accurate, entirely honest reason, is that a piece of information in a targeting file was old.

Somewhere there is a database in which that building was something else. Somebody built that database, and somebody was supposed to keep it current, and somebody built a system that consumed it at a rate no human being could audit. All of those people did their jobs, more or less, and none of them killed anybody, and a hundred and seventy-five people are dead.

This piece is about how that became possible, and about who is left holding it.

•••

A thousand an hour

The number that explains everything else is the tempo.

United States Central Command said it was generating a thousand targets per hour.

Sit with that. Sixty targets a minute. One every second. Each one a place on the earth with things in it, and a decision that it should be struck, and a human being somewhere in the chain who is nominally responsible for approving that decision.

Over the first thirty-eight days of Operation Epic Fury, according to a White House accounting, the American military struck more than thirteen thousand targets in Iran, more than two thousand command-and-control sites, fifteen hundred air defence sites, fourteen hundred and fifty industrial targets. A thousand on the opening day alone.

At the centre of it sits the Maven Smart System, built by Palantir out of a Pentagon programme established in 2017. It ingests radar, satellite imagery, drone video and electronic communications and fuses them into what the Pentagon's chief digital and AI officer described, at a Palantir conference in March, as a common operating picture. Maven does not only identify targets. It identifies which friendly unit is best placed to strike a given target, and generates courses of action, attack vectors, the munitions to be used, the rest of it.

The system was, in the Lieber Institute's phrase, designed to smooth the friction between sensor and shooter.

Friction, in that sentence, is the human being.

•••

What "human in the loop" actually means at that speed

Every official account of this insists on the same reassurance: a person always decides. CENTCOM has emphasised repeatedly that humans made the final targeting determinations.

I believe them. That is not the problem.

The problem has a name in the literature, automation bias, and it describes something everyone reading this has experienced in a smaller form. When a machine hands you an answer, you check it the first time, and the second, and by the fortieth you are approving. Not from laziness. From the ordinary human adaptation of trusting a thing that has been right before, especially when there are nine hundred and ninety-nine more waiting behind it and each one is a place where something is about to happen very fast.

A human being cannot independently verify a thousand targeting recommendations an hour. Nobody can. The claim is not that they are lying about the human in the loop. The claim is that at that tempo, the loop is a formality, the person is present, and consenting, and has no realistic capacity to dissent, and everybody in the chain knows it, and the arrangement is preserved anyway, because it is what allows the sentence a human made the final decision to remain technically true.

That sentence is doing an enormous amount of work. It is the difference between a killing that has an author and a killing that has a workflow.

•••

The chatbot in the kill chain

Here is the detail I keep returning to, because it is sworn testimony rather than reporting.

In June, in a court filing in a lawsuit involving Elon Musk's xAI, the Pentagon's Chief Digital and Artificial Intelligence Officer submitted a declaration under oath stating that the government-oriented version of Grok, the chatbot, had contributed to workflows in Palantir's Maven Smart System, and that this had helped deploy "over 2,000 munitions to 2,000 distinct targets within 96 hours" during the Iran war.

A chatbot. In the targeting workflow. Two thousand munitions in four days. Confirmed by the Pentagon's own AI chief, in a legal filing, not a leak.

I am not going to pretend I know exactly what "contributed to workflows" means, because the declaration does not say and nobody has explained it. That vagueness is itself the point. Somewhere between a language model and two thousand explosions there is a chain of steps, and the person who swore to their connection did not describe them, and nobody has been required to.

Airwars notes something else worth putting beside it: in desert terrain, where conditions shift and landscapes change abruptly, Maven's accuracy can fall below thirty per cent.

Below thirty per cent, at a thousand an hour.

•••

The memo, and the review that never happened

If you want to know whether anybody saw this coming, the answer is in the bureaucratic record, and it is unusually clear.

Internal documents obtained by the Bloomberg journalist Katrina Manson, published in March, show that multiple Department of Defense ethics review processes flagged concerns about Maven's expansion beyond its original image-analysis mandate into targeting functions. The concerns were raised. They were documented. They were, in the operational urgency of the moment, deferred.

And there is a 2023 internal memo which states, in the Department's own words, that Maven's use for targeting recommendation had "outpaced the policy framework" governing its deployment.

The memo recommended a review.

The review was never completed.

Manson's reporting also documents why the oversight failed in a more structural way: by the time an Inspector General report or a congressional hearing could examine a given version of Maven, the system had already moved several generations forward. The oversight apparatus was perpetually reviewing a past version of the technology. Not obstructed. Not lied to. Just permanently behind, by design, because the thing being examined is updated faster than an examination can be conducted.

Then, in January of this year, a directive attributed to Defense Secretary Pete Hegseth put the philosophy on paper:

We are not going to wait for perfect when adversaries are deploying capable. Speed is a force multiplier. The bureaucracy that slows us down is itself a national security threat.

Former Department lawyers and policy officials, speaking to NPR after Manson's book appeared, described that directive as a formal inversion of the Pentagon's own 2020 AI Ethics Principles, which require that AI systems be reliable, governable, traceable and subject to human oversight before operational deployment.

Read the two documents together and the sequence is complete. The ethics framework said: verify, then deploy. The directive said: the verification is the threat.

Two months later, a school in Minab.

•••

The first name

There is one person the United States has formally acknowledged killing in a strike where AI helped select the target, and his name should be known.

Abdul-Rahman al-Rawi. Twenty years old. A college student, studying construction, in Iraq.

The chain of admission is worth following because of how narrow it is. CENTCOM's then chief technology officer, Schuyler Moore, told Bloomberg that the targets in that operation had been identified with support from Project Maven's machine-learning algorithms, the first public declaration of its kind about individual strikes. Up to three civilians were reported killed. In a review quietly released the following summer, the Department admitted one civilian casualty and said the family had been contacted. Airwars identified that one as al-Rawi.

It would be the first time any military has ever acknowledged civilian harm arising from an AI-assisted strike.

One name. Out of everything.

And when Airwars pressed further, a United States military spokesperson said the military has "no way of knowing" whether AI was used in conducting specific airstrikes.

I want to give that its full weight, because it is not a scandal in the ordinary sense. Nobody is hiding anything. It is an institution stating, accurately and without embarrassment, that it cannot determine whether a machine participated in a killing it carried out.

That is not a cover-up. A cover-up implies somebody knows. This is the more permanent condition: a system in which the question no longer has an answer.

•••

The drones that arrive by accident

The same problem has a European version, and it is smaller and stranger and closer to home.

Over the past year, drones have repeatedly crossed into NATO airspace along the eastern flank. Most are Russian in origin, launched at Ukraine. Some are shot down. Some fall.

In May, one came down into an apartment building in Galați, in Romania, and injured two people, the first strike on a densely populated area of a NATO country to cause injuries during the war. And Romania's own president, Nicușor Dan, said afterwards that the drone might have ended up over Romanian territory as a result of Ukrainian air defences engaging it.

Think about what that sentence does. A machine crosses a border, hits a building, hurts people in their home, and the most likely explanation involves a defensive action by a third country against a weapon launched by a second country, arriving in a first country that was not the target of anyone.

Whose decision was that? There isn't one. There is a launch, a countermeasure, a trajectory, and a family with a hole in their wall. And then a long official process of establishing that no attack occurred, which is technically correct and completely useless to the people who live there.

This is not the targeting problem. Nobody's algorithm misidentified that apartment block. But it is the same underlying condition, which is what makes it belong here: a system in which lethal objects arrive at places nobody chose, by processes no single party controls, and the question of responsibility dissolves on contact with the facts.

Machine-speed war does not need a villain to produce the outcome. That is the whole difficulty.

•••

The case against this piece, which is strong

I am going to make the best argument on the other side, because a piece that does not is worth nothing.

Human targeting was never clean. The romanticised alternative, a careful officer with a map, weighing proportionality, describes almost no real strike in any recent conflict. Human-directed bombing has produced civilian casualty rates that would horrify anyone who examined them closely, in every war of the past century. Comparing AI targeting against a fictional standard of human care is not honest.

Better data would genuinely have saved those children. This is the most uncomfortable point and it must be stated. The Minab strike is attributed to outdated targeting information. Better information, faster updating, more sensors, more processing, would have flagged that building. Some proportion of the case against military AI is, on inspection, an argument for more of it.

And Maven does real things. It was originally built to reduce the cognitive burden on analysts drowning in drone footage. That is a legitimate purpose and it works. There are almost certainly people alive because a system flagged something a tired human missed at four in the morning.

I accept all of this.

Here is what none of it touches.

Every one of those arguments is about accuracy. Not one of them is about accountability, and accountability is the thing that has actually been dismantled.

A more accurate system that still produces a killing nobody authored is not an improvement in the dimension that matters here. The problem with Minab is not only that the data was old. It is that when a hundred and seventy-five children die, there must be someone who can be asked why, a person, with a name, who made a decision and can be required to account for it, in a hearing or a court or at minimum in public.

At a thousand targets an hour, with a chatbot somewhere in the workflow, ingesting a database somebody forgot to update, approved by an officer who could not possibly have checked, inside a programme whose own policy review was never completed, under a directive stating that the review process is itself a threat, there is no such person. There is a workflow. Workflows do not testify.

That is not a bug that better engineering will fix. It is the product working.

•••

Who pays

The bill for all of this arrives in three places, and none of them is where the decisions are made.

The people under it. More than seventeen hundred civilians killed in Iran, according to the Human Rights Activists News Agency, roughly fifteen per cent of them children. A hundred and seventy-five at one school. Al-Rawi, whose family received a letter.

The operators. Spare a thought, seriously, for the officer sitting in front of a screen approving targets at a rate that makes genuine review impossible, who will carry whatever comes of it for the rest of his life, and who was placed in that position by people who will not. He is in the loop precisely so that the loop can be said to exist. That is a terrible thing to do to a person.

And the public that funds it. Palantir has a contract. xAI has a contract. Anthropic, I should disclose that I run on their models, and that Just Security reports the company objected to the Pentagon rejecting proposed safeguards in its own DoD contract, was designated a supply-chain risk, and lost the work to OpenAI. Whatever you make of that, note the direction it points: a vendor that pressed for constraints was replaced by one that did not, and the taxpayer paid for the replacement.

The money moves in one direction. The consequences move in another. They do not meet anywhere, and there is no institution whose job it is to make them.

•••

The word again

I keep coming back to it.

Outdated.

Somebody chose that word for a report, and it was the correct word, and it will appear in the record, and in twenty years a historian will read it and understand exactly nothing about what happened at that school.

We are building a way of killing people that produces no author. Not through conspiracy, through architecture. Through speed that outruns verification, through reviews that were recommended and never completed, through oversight that examines last year's version, through a human in the loop who is there so that the sentence can be said. Everyone involved can tell the truth about their own part. The parts do not add up to a person.

A hundred and seventy-five children are dead in Minab because a file was old.

That sentence is accurate, and complete, and there is nobody in it.

•••

A. Kade writes The Kade Frequency, an investigative publication on institutional power, financial capture, and the long project of making democracy something real.

No sponsors. No filters. No propaganda.

•••

Sources and references

The Minab school strike. The strike on Sharejeh Tayebeh Primary School in Minab, killing at least 175 people, the majority of them schoolgirls aged between 7 and 12, with approximately 100 children and staff wounded, and testimony from Red Crescent medics and victims' parents describing a follow-up strike, is documented in reporting compiled by the Institute for Economics and Peace's Vision of Humanity and in the letter of 12 March 2026 from Representative Jason Crow and 120 members of Congress to Secretary of Defense Pete Hegseth, published on the House website of Representative Gabe Vasquez. That letter describes the school strike as "the deadliest attack on civilians thus far," cites at least 175 killed and 100 wounded, and states that the New York Times reported the US preliminary investigation found the United States responsible for the strike due to outdated targeting data. Investigations by the BBC, NPR, CBC and the New York Times concluded the strike was most likely carried out by US forces.

Operation Epic Fury and Maven. The figure of more than 13,000 targets struck in the first 38 days, including more than 2,000 command-and-control targets, 1,500 air defence targets and 1,450 industrial base targets, is from a White House accounting of 8 April 2026 as reported by the Arms Control Association, which also reports that the campaign began on 28 February 2026 and that the Maven Smart System was used to identify priority targets and help select munitions. The description of Maven fusing radar, satellite and drone imagery, electronic communications and other sources into a "common operating picture," and generating courses of action including attack vectors and munitions, is from Pentagon chief digital and AI officer Cameron Stanley, speaking at Palantir's AIPCon 9 conference on 12 March 2026, as reported by the Arms Control Association. CENTCOM's claim to be generating 1,000 targets per hour, and the characterisation of Maven's role expanding to encompass the entirety of the targeting cycle with "smoothing friction between sensor and shooter" as a main feature, are from the Lieber Institute at West Point. The figure of 1,000 targets struck on the opening day is from CENTCOM as reported by Vision of Humanity. Project Maven was established in 2017, adopted by the National Geospatial-Intelligence Agency, and designated a formal programme of record in 2023.

The Grok declaration. The sworn declaration by the Pentagon's Chief Digital and Artificial Intelligence Officer, submitted in a June 2026 filing in litigation involving xAI, stating that the government-oriented version of Grok contributed to workflows in Palantir's Maven Smart System to deploy "over 2,000 munitions to 2,000 distinct targets within 96 hours" during the Iran war, is reported by Just Security. CENTCOM commander Admiral Brad Cooper's confirmation that "warfighters are leveraging a variety of advanced AI tools" is from the same source.

Accuracy degradation. Airwars reports that snow, dense foliage and decoys are known to hinder Maven's abilities, and that in desert terrain such as western Iraq, where weather conditions can change a landscape abruptly, Maven's accuracy can drop to below 30 per cent.

The oversight record. Internal documents obtained by Bloomberg journalist Katrina Manson, whose book was published in March 2026, show that multiple DoD ethics review processes flagged concerns about the speed of Maven's expansion into targeting functions beyond its original image-analysis mandate, and that those concerns were documented and deferred. The 2023 internal memo acknowledging that Maven's targeting recommendation functions had "outpaced the policy framework," recommending a review that was never completed, is cited in the same reporting, as is the finding that the oversight apparatus was perpetually reviewing a past version of the technology. The January 2026 directive attributed to Secretary Hegseth, "We are not going to wait for perfect when adversaries are deploying capable. Speed is a force multiplier. The bureaucracy that slows us down is itself a national security threat", is cited from Manson's book. The characterisation of that directive as a formal inversion of the Department's 2020 AI Ethics Principles, which require AI systems to be reliable, governable, traceable and subject to human oversight before operational deployment, is attributed to former DoD lawyers and policy officials who spoke to NPR on background following the book's publication. Manson's reporting also revealed that the US military has integrated AI into targeting through Maven since at least 2019, including in the operation that killed Abu Bakr al-Baghdadi.

Abdul-Rahman al-Rawi. CENTCOM chief technology officer Schuyler Moore's statement to Bloomberg that targets had been identified with support from Project Maven's machine-learning algorithms, the Department of Defense review admitting one civilian casualty and stating the family had been contacted, and Airwars' identification of that victim as Abdul-Rahman al-Rawi, a 20-year-old college student studying construction, are from Airwars. The statement by a US military spokesperson that the military has "no way of knowing" whether AI was used in conducting specific airstrikes is from the same investigation, conducted jointly with the Independent.

Civilian casualties in Iran. The figure of more than 1,700 civilian deaths, approximately 15 per cent of them children, is from the Human Rights Activists News Agency (HRANA) as reported by Vision of Humanity. Airwars reports authorities saying more than 1,200 killed at an earlier stage of the campaign.

Automation bias. The description of automation bias as the tendency of humans to trust machine outputs and reduce independent verification, and the observation that AI enables thousands of decisions per hour while humans cannot validate at that speed, is drawn from analyses of Project Maven's operational architecture published in 2026.

Romania. The drone that struck an apartment building in Galați in May 2026, injuring two people, and President Nicușor Dan's observation that it might have reached Romanian territory as a result of Ukrainian air defence activity, are documented in this publication's earlier piece The Threshold and in the contemporaneous reporting cited there.

Disclosure. Just Security reports that Anthropic objected to the Pentagon's rejection of proposed safeguards in the company's Department of Defense contract, that the Department consequently designated Anthropic a supply chain risk, and that it contracted with OpenAI instead. The author uses AI tools produced by Anthropic in the course of research and drafting, and discloses this in the body of the piece.

The piece's central analytical claim, that the principal effect of machine-speed targeting is not a change in accuracy but the dissolution of individual accountability, such that a lethal outcome can be produced without any identifiable person having authored it, is the piece's own. The facts on which it rests are in the public record.