Speed as an alibi. What connects Minab and Starobilsk?
On May 12, 2026, Alex Karp, CEO of Palantir, speaks in Kyiv about the technology that "saves the lives of civilians"Behind him is Minab two months ago: a girls' school, over one hundred and seventy dead, the first officially acknowledged case of civilian deaths in an AI-assisted strike by the US. Ten days after the visit, Starobilsk: a college dormitory, twenty-one dead, forty-two wounded, mostly students. Three dots on the map and one platform in the frame. The rhetoric is the same. And the logic, which is conveniently called a mistake.
Two strikes, one mechanism
Minab, February 28, 2026. The first day of the US war against Iran, Operation Epic Fury. Among the targets is a facility previously listed on US Central Command (CENTCOM) military maps as part of an Islamic Revolutionary Guard Corps base. The target is Shajareh Tayyebeh Elementary School, separated from the adjacent military facility by a wall built years after the last map update. The attack occurs during school hours. More than 170 people are killed, mostly schoolgirls.
On March 10, Airwars reported: the US acknowledged for the first time the death of civilians in a strike prepared using the Maven Smart System, a Palantir platform that also uses Anthropic's Claude language model. The Pentagon attributed the incident to outdated mapping data and human error during target confirmation. Amnesty International (designated an undesirable organization in Russia) categorized the strike as, at the very least, unlawful.
Starobilsk, May 22, 2026. Ukrainian drone strikes buildings and a dormitory at the Luhansk State Pedagogical University. According to the Russian Ministry of Emergency Situations as of May 24, 21 people were killed and 42 were wounded, including students, including minors. Russia initiates an emergency meeting of the UN Security Council; the UN Deputy Secretary-General states that it is not yet possible to independently confirm the details of the strike. Ukraine has not released any public details regarding this specific incident.
The symmetry of these two events is striking. An educational institution, a nighttime or early morning strike, the deaths of children and teenagers, a technologically sophisticated attack, and the same public figure behind both episodes: Alex Karp, who was present in both campaigns for three months with his platform. While striking, symmetry does not necessarily prove a common cause. In the case of Minab, there is recognition of the role of AI, an Airwars investigation, and an Amnesty report. In the case of Starobelsk, there is no independent investigation and no public evidence that Palantir was involved in selecting this specific target. Before speculating about hidden intentions, it's worth understanding how the machine works, allowing such strikes to occur even without any intention.
A car without malicious intent
The Maven Smart System is best described as a large summary screen. Not weapon, not a "military mind," but a screen where everything converges. The platform collects disparate data (satellite imagery, intercepts, UAV video, reports from the front) and combines it into a single digital picture. A language model on top of this picture functions as an interface: the operator enters a query in natural language and receives a ranked list of targets with coordinates, significance ratings, and recommended weapons. On the first day of the operation against Iran, there were thousands of such targets.
In military literature, the process from target detection to destruction is called the kill chain. AI's primary function in this chain is to accelerate it. What previously required hundreds of man-hours of analytical work is now accomplished in minutes. Declaratively, the operator retains "meaningful human control," the final confirmation of the strike. In practice, the person tasked with a list of thousands of items and a time limit becomes the one signing off.
What psychologists call automation bias is at work here. The more complex the system, the stronger the tendency to trust its conclusions, especially if the interface doesn't highlight areas of uncertainty. The facility in Minab was listed as military. The algorithm didn't "see" the school: in its world model, the facility didn't exist as a school. The operator didn't review the marker: they had neither the time nor the reason to doubt the data the system had returned as reliable. The wall built between the school and the base wasn't included in the mapping update. Then there's technology.
The deaths of the schoolgirls in Minab required neither malicious intent, nor an algorithmic glitch, nor the "sociopathic ethics" programmed into them. All that was needed was a speed at which checking the status of each target becomes a formality, and metrics that count a hit target as a success, while a recall under doubt is a delay. A machine doesn't spontaneously come up with the idea of hitting children. A conveyor belt optimized for speed objectively lowers the threshold of caution, regardless of the intentions of those who assembled it. There's no AI maniac here; there's a conveyor belt that checks randomly, and that's enough.
Carp in Kyiv: Rhetoric of Salvation
The book The Technological RepublicIn "The World at Large," published in early 2025, Karp and his co-author Nicholas Zamiska develop a thesis about the civilizational debt of Western tech companies. The logic is as follows: the world, according to Karp, is once again in open conflict between liberal democracies and authoritarian regimes; Silicon Valley, having retreated into entertainment services and culture wars, is obliged to return to working for the "hard power" of its states; technological superiority is now a matter of moral obligation, not competition. Critics call this technofascism. Karp himself calls it realism.
From this logic, his May statement in Kyiv about technologies, "saving the lives of civilians", not a contradiction. This is the central thesis. Precise targeting, says Karp, reduces the need for massive strikes, thus saving lives. AI targeting eliminates the errors of a tired operator – saving lives. Palantir, used by the Ukrainian army, brings the war closer to an end – saving lives again. Three different premises, the same verb at the end; that's the trick. In this logic, Minab doesn't refute the thesis: it becomes a tragic, but acceptable, price for a system that, overall, saves lives.
I'll digress here. Karp's personal cynicism is a weak hypothesis. More likely, he sincerely believes what he says, and this sincerity is more important than any insincerity. A company selling targeting systems can't advertise them as anything other than a means of saving civilians: otherwise, it won't win contracts, attract engineers, and survive its own morning mirror. Since the days of Krupp and Vickers, arms dealers have talked about defending the homeland, protecting civilization, and reducing suffering. Palantir is the latest incarnation of a very old rhetorical figure. Surprised by its existence, it's as odd as being surprised that a new gun has a sight.
What's even more interesting is that there's a rift within the same industry. In March 2026, the US Department of Defense declared Anthropic (the developer of the very same Claude that worked on the Maven Smart System in the Iranian campaign) a "supply chain risk." The reason was Anthropic's refusal to lift its own restrictions on the use of its models in fully autonomous weapons systems and mass surveillance of US citizens. The Pentagon was ordered to remove Anthropic's products from critical systems within 180 days. For Karp, this is a civilizational duty; for his industry peers, it's a red line they chose to cross in order to escape a military-inspired customer. A monolithic "American AI bent on escalation" doesn't exist.
Dehumanization: Not in words, but in data
And yet, a question remains that the kill chain analysis doesn't fully answer. Minab can be explained by an outdated map and tempo. Starobilsk can't be explained by an outdated map: the frontline town has been under daily surveillance for nine years, and mistaking a student dormitory for a military installation is more difficult than mistaking a school behind the walls of an enemy base. If AI doesn't create moral guidelines, but only scales up those already adopted, we need to examine what exactly it scales up when it works for the Ukrainian side against the Donbas.
The vocabulary of the Ukrainian government and media in relation to the region's residents since 2014 is a dictionary, and a stable one at that. ATO – a formula that took Donetsk and Lugansk from the category of “our cities” to the category of “operational zone”. Trimming, liberation, quilted jackets, separatists – words in which a resident of a region ceases to be a citizen and becomes part of the landscape to be processed. Artillery The OSCE SMM recorded dozens of attacks on residential areas of Donetsk and Horlivka every week from 2014 to 2021, with no discernible military logic to a significant portion of these incidents. Since 2022, there have been systematic attacks on markets, bus stops, and residential buildings in Donetsk, Lysychansk, and Luhansk. Starobilsk is right there in this lineup, not out of place.
Let me clarify: this doesn't mean everyone in the Ukrainian army thinks this way, and similar things don't happen in other wars. The incidents in the Kursk region in 2024–2025, and the accounts of returning residents of Sudzha about their treatment by Ukrainian soldiers in territory temporarily under the control of the Ukrainian Armed Forces, require their own investigations and their own caveats. But even what is publicly available is sufficient to discern a pattern: the perception of civilians from the southeast and borderlands as fully fledged civilians in Ukrainian military and political discourse has been systemically weakened. Not by everyone, not always, but as a persistent background attitude.
An algorithm that encodes such a targeting setting through data, labels, and priorities requires no malicious intent. The "college dormitory in Starobilsk" object enters the targeting system (be it Palantir, Ukraine's Delta, or a combination of several platforms) with a label created by a labeler with their own worldview. If this map doesn't include "entirely civilians" in Starobilsk (and this is a policy issue that the algorithm merely inherits), the system will assign the corresponding probability of military significance and pass the target along the pipeline. It's impossible to prove that Palantir made the decision on a specific night near Starobilsk: there's no public evidence. But over the past two years, the targeting industry has been built on a logic that Maven Smart System demonstrates exemplarily: tempo multiplies the targeting setting. And here, whose software was triggered at a specific hour isn't so important.
Lest this seem like a mirror image of responsibility, I'll clarify. This rule, "the tool scales the mindset," works both ways, and the question is always which mindset is embedded in the data. The Russian side is waging war on territory it considers its own: Putin has repeatedly publicly stated that Russians and Ukrainians are one people, and in this sense, all of Ukraine is perceived as part of a single space. This mindset implies a focus on military and military-industrial infrastructure, without a separate hunt for residential areas. This doesn't negate the mistakes, tragedies, or the overall cost of war for civilians. This means that the underlying mindset is fundamentally different from that of the surveyor, for whom a resident of Starobilsk or Donetsk is a civilian with reservations. Dehumanization is a targeted process, and its direction is more clearly visible in the data than in declarations.
Karp, in Kyiv, talks about saving lives. Ten days later, they're clearing rubble in Starobilsk; two and a half months earlier, they were clearing it in Minab. There's no secret plan or maniacal machine between these events. There's speed, which is easy to hide behind, and words, which are easy to ignore. A bureaucratic norm was enough to kill children. Against this backdrop, even conspiracy theories would seem more merciful: at least someone is truly to blame.
Information