Pentagon says AI overreliance contributed to deadly Iran school strike

On February 28, two US Tomahawk missiles struck Shajarah Tayyebeh Elementary School in the southern Iranian city of Minab, a 260-pupil school built on land that had been part of an Islamic Revolutionary Guard Corps (IRGC) naval compound. The strikes killed more than 150 people, including at least 123 children, making it the deadliest American military targeting error involving children this century. According to Bloomberg's reporting, based on interviews with more than two dozen officials involved in or briefed on the Pentagon's internal investigation, the disaster did not result from one catastrophic decision but from an accumulation of smaller ones: flawed intelligence, outdated satellite imagery, cuts to civilian-harm-mitigation staff and overreliance on artificial intelligence software.

The Minab site had been catalogued years earlier as a military compound, a classification that was never updated even though the location had visibly changed. The New York Times reported in June that targeting officials relied on satellite imagery that had not been updated in seven years and did not show a school on the site; commercial satellite images show walls and entrances separating the school from the base were built by 2017, with a soccer pitch and playground markings visible by 2018. One US intelligence analyst had detected the changes as early as 2019, Bloomberg reported in June, but logged the observation in a system not connected to the primary targeting database. Investigators have focused on the Defense Intelligence Agency's failure to reclassify the site.

Inside US Central Command, which carried out the strike, some personnel leaned too heavily on Palantir Technologies' Maven Smart System, the AI-enabled software Centcom uses to fuse more than 150 data inputs and coordinate targeting, officials told Bloomberg. Work that previously took hours, compiling and cross-checking proposed target lists, was compressed into several minutes using Maven ahead of the war. Some Centcom personnel expected the tool to flag stale intelligence or inconsistencies on its own; officials could not say why they held that expectation, and it did not happen for the Minab site. A Palantir spokesperson said the company "is not responsible for the underlying data nor identifying intelligence deficiencies" and that there is no evidence its software was at fault; two people familiar with Palantir's Pentagon contracts said the government retains primary responsibility for data quality, though it is not uncommon for military users of tools like Maven to develop expectations of the software that differ from the contract terms. After the strike, Palantir added a new capability to Maven that re-reviews underlying intelligence to flag inconsistencies a human reviewer might miss, according to a person familiar with the matter.

The other major failure investigators identified was staffing. Defense Secretary Pete Hegseth had dismantled most of the Pentagon's civilian-harm-mitigation (CHM) units before the war, cutting headcount across those teams by roughly 90%, to fewer than 20 staff members; Centcom's own CHM team shrank from 10 people to one. No CHM reviewer examined the Minab site before the strike, a result both of the staffing cuts and of a decision to keep the group out of planning, officials said, even though such review had become routine, if not mandatory, in recent years. Hegseth has described the cuts as a break from "Biden-era, non-lethal programs"; the Biden administration had built up the CHM capability after a 2021 US drone strike killed 10 civilians in Kabul based on a misidentified aid worker. "America, regardless of what so-called international institutions say, is unleashing the most lethal and precise air power campaign in history," Hegseth told reporters on March 2, days after the strike. He added: "No stupid rules of engagement, no nation-building quagmire, no democracy building exercise, no politically correct wars."

The strike came amid an extraordinary tempo: more than 1,000 Iranian targets were hit in the first 24 hours of the war alone, after diplomatic talks between the US and Iran in Geneva collapsed on February 26. Vetting that had taken weeks or months in earlier conflicts, matching sites against target libraries built up over decades, was compressed into days, and more than a thousand potential targets had to be reassessed and approved. Roughly three dozen people worked the Minab site along the targeting "kill chain," most of them more than 7,000 miles away in Tampa, Florida. Laurie Blank, a special counsel at the Pentagon's general counsel's office from 2022 to 2024, said "Haste can lead to errors," distinguishing rushed strikes on preplanned targets, which happened here, from split-second decisions made in response to an ongoing attack. Investigators who reviewed the satellite imagery, intelligence databases, chat logs and sign-offs concluded there was high confidence throughout the chain that the site was still an IRGC facility. "Insufficient intel wasn't really viewed as a concern before the war," one Pentagon official involved in the probe told Bloomberg. "We were confident in our intel. Minab proved you should never be too confident."

A UN Independent International Fact-Finding Mission on Iran concluded this week that the US had "failed in its obligation to do everything feasible to verify" that the school was a military target and that the failure "went beyond negligence," and separately found reasonable grounds that the Minab strike and another US attack the same day amounted to war crimes. The US has not publicly accepted responsibility; President Trump told Fox News on July 14, "I don't think anybody's going to ever be able to say what happened there," and a senior administration official said only that "the United States does not target civilians." Admiral Brad Cooper, who heads Centcom and ordered the internal investigation, told Congress in May he was committed to transparency, but the full investigation report, underway since March and described by officials as all-but-complete for months, still has not been released. More than 120 House Democrats wrote to Hegseth in March asking about AI's role in the strike without a public response. Senator Jack Reed accused the Pentagon of withholding the results; his Republican counterpart Thom Tillis called for congressional briefings, and Senator Richard Blumenthal, who sits with them on the Senate Armed Services Committee, said, "The administration should have been forthcoming within hours or days and here we are months later and nothing has been provided. It is unforgivable."

Minab is one entry in a wider pattern of US targeting failures involving flawed intelligence, including the 1999 bombing of the Chinese Embassy in Belgrade (three killed, 27 injured), a 2015 airstrike on a Doctors Without Borders hospital in Kunduz, Afghanistan (42 killed), and an April 2025 strike on the Ras Ina port in Yemen (at least 80 civilian deaths), which came shortly after the Pentagon's civilian-harm-mitigation teams were disbanded. According to Human Rights Activists in Iran, at least 1,701 civilians were killed in the war's first 39 days, an average the group itself calls a low estimate; its deputy director, Skylar Thompson, said the question is "how so many civilians were killed amid an unprecedented tempo of strikes over just a matter of weeks." A September 1 strike on Kuhestak, near Minab, killed at least five more people at a wedding party; Vice President JD Vance said the incident is under investigation.

On the day of the attack, teachers had begun sending children home early after reports of strikes on Iran; two missiles hit minutes apart, collapsing part of the school roof onto a prayer room where students had taken shelter. According to accounts gathered by human rights investigators and the watchdog group Airwars, parents identified their children afterward by fragments of clothing; one father recognized his 7-year-old son, Arash Gol-Azin, only by his sock. Payam Akhavan, a lawyer who has practiced before the International Criminal Court, said investigators need to establish what steps US officials took to verify the target: "The Geneva Convention says you must take all feasible measures to determine that this is, in fact, a military objective. But if all feasible measures were not taken, then that raises a question of criminal responsibility."

Key facts

Two Tomahawk missiles hit Shajarah Tayyebeh Elementary School in Minab, Iran on February 28, killing more than 150 people, including at least 123 children, the deadliest US targeting error involving children this century.
The site's classification as an IRGC military facility was never updated: targeting officials used satellite imagery seven years out of date, even though one analyst had flagged changes at the site as early as 2019.
Defense Secretary Pete Hegseth's Pentagon cut civilian-harm-mitigation staff by roughly 90% (Centcom's own team fell from 10 people to 1), and no such reviewer checked the Minab site before the strike.
Centcom personnel leaned heavily on Palantir's Maven Smart System, which fuses more than 150 data inputs and compressed target vetting from hours to minutes; Palantir says it is not responsible for the underlying intelligence and denies its software was at fault.
A UN fact-finding mission found the US failed to take feasible measures to verify the target and said the failure "went beyond negligence"; more than 120 House Democrats and senators from both parties have pressed for transparency the Pentagon has not delivered months later.

Why it matters

This is one of the most detailed public accounts of how a commercial AI targeting tool, Palantir's Maven Smart System, fed into a specific fatal military decision, with Pentagon officials themselves describing overreliance on it as one of several contributing causes. The strike is the deadliest documented American targeting error involving children this century, and it happened after human oversight was deliberately thinned: a near 90% cut to civilian-harm-mitigation staff meant no dedicated reviewer looked at the site, while an AI system that fuses more than 150 data feeds compressed vetting from hours to minutes. The case sits squarely in the debate over how much speed and autonomy a government should hand to AI-assisted decision systems in life-and-death contexts, and over who is accountable when the system does not catch what a slower human process might have.

Who it affects

Palantir, whose defense-AI business and reputation are directly implicated even as it denies its software was at fault; Pentagon and Centcom leadership, including Hegseth and Admiral Cooper, who face congressional pressure from both parties; the survivors and families in Minab, and Iranian officials who have used the strike politically; US lawmakers weighing future defense AI budgets and oversight ahead of the November elections; and, more broadly, any government or organization building AI-assisted targeting, triage or other high-consequence decision tools, which now has a concrete case study of what happens when speed, thinned human review and confidence in a tool combine.

How to use it

There is no product here to adopt, but the episode is a concrete lesson for anyone building or buying AI decision-support tools for high-stakes, time-pressured settings: do not assume a tool will catch stale or inconsistent data on its own unless that capability is explicitly built, tested and made mandatory, and do not let a compressed AI-assisted review replace a dedicated human harm-mitigation step just because that step has become routine rather than required. Palantir's post-strike addition to Maven, a feature that specifically re-reviews intelligence to flag disqualifying factors and inconsistencies a human reviewer may miss, is the kind of safeguard officials say they had expected the system to already provide before the strike. Building or mandating that check up front, rather than adding it after a failure, is the actionable takeaway.

How solid is it

The account rests on Bloomberg's interviews with more than two dozen current and former officials directly involved in or briefed on the Pentagon's internal investigation, plus named public sources: a UN Independent International Fact-Finding Mission report released this week, New York Times reporting in June on the outdated satellite imagery, Bloomberg's own earlier reporting on the 2019 analyst flag, a Pentagon spokesperson's on-record confirmation that the incident remains under investigation, and on-record statements from Palantir, senators and a senior administration official. The Pentagon's own full investigation report, underway since March, has not been released, so its formal conclusions are not yet public, and most officials who described the probe to Bloomberg spoke on condition of anonymity given the sensitivity of the material, which limits independent verification of specifics such as exact staffing numbers or the minute-by-minute sequence inside the kill chain.

Risks and caveats

Palantir disputes that its software was at fault and says responsibility for data quality lies with the government, so the causal story is contested at exactly the point that names the company. The source does not identify any individual analyst, operator or commander who made the final targeting call, does not give a release date for the Pentagon's full report, and does not say whether Maven's targeting role has since been restricted beyond the added re-review feature; those remain open questions, not settled facts. The war's broader civilian toll and the political use of the Minab strike, by Iranian officials and by US lawmakers ahead of a November election, mean the incident is being read through partisan and diplomatic lenses on multiple sides, on top of its substance as a case of AI-related failure.

“Insufficient intel wasn't really viewed as a concern before the war. We were confident in our intel. Minab proved you should never be too confident.”

— a Pentagon official involved in the internal investigation, quoted by Bloomberg