Fridays, 2.30 pm (Lisbon) · Artificial intelligence on NOW →Fridays, 2.30 pm · AI on NOW →
Miguel Sousa SoaresMSS
← Writing
LinkedIn · Essay · 9 April 2026

The Machine That Decides Who Dies

A primary school appeared on a target list. Europe has already adopted the technological logic that made this possible.

On the morning of 28 February 2026, a Saturday, around 180 children entered the Shajareh Tayyebeh girls’ primary school in the Iranian city of Minab. Saturday is a school day in Iran. The youngest were seven years old.

Satellite images analysed by Al Jazeera and BBC Verify show the building intact at around 10:23 a.m. local time. Shortly after 10:45, it was gone. The school was hit by several missiles in succession. The roof collapsed on the pupils. More than 165 people were killed. More than a hundred of them were children.

The school stood less than a hundred metres from a naval compound belonging to the Islamic Revolutionary Guard Corps. A wall built between 2013 and 2016 physically separated the two sites; they had separate entrances, and the watchtowers had been removed. The school had a website, a social-media presence and open enrolment. Anyone with internet access could verify what operated there. And yet those coordinates found their way onto a target list.

To plan Operation Epic Fury, the US armed forces used the Maven Smart System, a platform developed by Palantir under a $1.3 billion Pentagon contract. Maven fuses data from more than 150 sources in a single interface. It classifies targets, recommends munitions and generates automated legal justifications for each strike. Until recently, it ran on artificial-intelligence models from Anthropic, the company behind Claude.

In the first eleven days of operations in Iran, the United States carried out 5,500 strikes. A workload that previously required around 2,000 intelligence analysts was handled by teams of twenty. Cameron Stanley, the Pentagon’s head of AI, described Maven as revolutionary: a process that used to take weeks could now be measured in seconds.

This is precisely where the problem begins.

According to Pentagon data reported by Bloomberg in 2024, Maven correctly identifies objects in about 60% of cases. Sixty per cent. Trained human analysts are right 84% of the time.

There is an even more troubling figure. In 2021, an experimental US Air Force targeting AI achieved 25% accuracy in real-world conditions while rating its own confidence at 90%. Twenty-five per cent accuracy. Ninety per cent confidence. It was wrong three times out of four, and thought it was almost always right.

So let us ask the question that really matters. If Maven generated hundreds of strike coordinates in the first twenty-four hours of the campaign against Iran, and its real-world reliability is lower than that of a trained analyst, how many of those coordinates were wrong? How many pointed to schools, hospitals, pharmacies or homes? The Pentagon did not answer. The investigation is ongoing.

The preliminary investigation into the Minab strike points to outdated data, not an autonomous algorithm failure. Former military officials told Semafor that people fed the system obsolete information: years after a wall had separated the building from the military compound and it had been converted into a school, it was still listed as part of the military site.

That does not make the case less serious. It makes it more revealing.

The promise of military AI was precisely to reduce this kind of error: process more sources, cross-check more data, detect what the human eye misses. Maven receives updated satellite imagery. It receives real-time drone feeds. The Minab school had a verifiable online presence. And yet the error was not corrected. It was accelerated.

The problem was not just the algorithm. It was the architecture of trust around it. When 2,000 analysts are reduced to twenty, more than personnel are lost. Friction is lost. The analyst who looks at the screen and says, “Wait, there was a school here” is lost. The second pair of eyes asking for confirmation is lost. The hesitation that, in a decision chain compressed into seconds, was the last human safeguard is lost.

Efficiency removed exactly what might have stopped the error.

There is a political dimension to this story that should not be ignored. Anthropic, whose models were integrated into Maven, refused to allow Claude to be used in autonomous lethal weapons or the mass surveillance of US citizens. The Pentagon demanded unrestricted access. Negotiations collapsed.

The US government designated the company a supply-chain risk, a label normally reserved for entities linked to foreign adversaries. A federal judge, Rita Lin, provisionally blocked the measure. Anthropic therefore won an initial court victory, but not a final one. Shortly afterwards, OpenAI signed its own agreement with the Pentagon. Sam Altman said the agreement included safeguards. Critics noted that the contractual language was substantially more flexible than the red lines Anthropic had drawn.

The meaning is clear. The only major AI company that refused to remove restrictions on more sensitive military uses was pushed out. The machine continued, with another supplier.

And where is Europe in all this?

In March 2025, NATO acquired the Maven Smart System. The platform is installed at Allied Command Operations headquarters in Mons, Belgium, with American technology, American models and American cloud infrastructure. European armed forces have integrated a life-and-death decision technology they did not build, do not control, cannot audit and whose inner workings are entirely opaque to them. They did so without a single public debate, meaningful parliamentary scrutiny or a single line of code of their own.

The problem becomes more serious when we read what US strategists themselves acknowledge. According to Benjamin Jensen and Yasir Atalan of the Center for Strategic and International Studies, the Pentagon lacks a systematic framework for evaluating AI in operational settings. Only around 2% of global AI evaluation efforts are devoted to defence. Even those efforts focus mainly on cybersecurity, leaving out functions such as targeting, logistics, intelligence fusion, command and control.

If the United States, which develops, funds and operates these platforms, still cannot robustly measure their reliability in combat, what capacity does Europe have to do so with systems that are not even its own?

It is like flying an aircraft using instruments that have never been properly tested by the people in the cockpit. And when the aircraft crashes, it is not the instrument manufacturer that goes to court.

Meanwhile, Maven has been formalised as an official Pentagon acquisition programme, with a $1.3 billion contract and guaranteed multi-year funding. It has more than 20,000 active users. The platform has expanded. European allies have adopted it without serious debate about what it does, how it decides or what margin of error it has.

In Minab, more than 165 people died in a school with a website, an online presence and children in its classrooms. The system that helped plan the strike is less accurate than a human analyst. And the only major AI company that said no was pushed out of the process.

Perhaps the most troubling question is no longer whether the machine made a mistake in Minab.

Perhaps it is this: when the decision to kill moves at the speed of software, how much room is there, in practice, for a human being to stop it?

Methodological note

The data and facts cited in this article draw on reporting and investigations by The New York Times, BBC Verify, NPR, Reuters, Al Jazeera, The Washington Post and Human Rights Watch on the attack on the Minab school on 28 February 2026; Military Times, The Register, Bloomberg and Defense One on the Maven Smart System and its operational reliability; Semafor and NBC News on the ongoing investigation and human versus algorithmic responsibility; CNBC, CNN, Axios and NPR on the dispute between Anthropic and the Pentagon; The Guardian, Politico and OpenAI on the agreement with OpenAI; SHAPE/NATO and NCIA on NATO’s acquisition of the platform; DefenseScoop, Tom’s Hardware and other market sources on the programme’s formalisation; Benjamin Jensen and Yasir Atalan of the Center for Strategic and International Studies on the AI benchmarking gap in military evaluation; and the Palantir Blog on the platform’s demonstration in the NATO context. The analysis and opinions expressed are the author’s alone.

Translation of the text in the author’s archive. Original publication: A Máquina que Decide Quem Morre · LinkedIn · Essay · 9 April 2026.
Original bibliographic details (Portuguese)

Ensaio de Miguel Sousa Soares, publicado originalmente no LinkedIn a 9 de abril de 2026. Esta página é o arquivo pessoal do autor.

Miguel Sousa Soares is an economist and analyst of geoeconomics and artificial intelligence, with more than twenty years’ experience in strategy and corporate finance. He writes for Visão, Jornal de Notícias, Expresso and Correio da Manhã. He comments on artificial intelligence on NOW on Fridays.