On February 28, two Tomahawk missiles slammed into the Shajarah Tayyebeh Elementary School in Minab, southern Iran. The blasts killed more than 150 people. At least 123 of them were children.
The strike came in the opening hours of a sweeping U.S. and Israeli military campaign against Iran. Planners had targeted what they believed was an Islamic Revolutionary Guard Corps naval facility next to the school. They got the wrong building. And now a still-unreleased internal Pentagon review lays bare a chain of errors that led there.
According to officials familiar with the investigation, U.S. personnel at Central Command placed too much faith in an artificial intelligence platform built by Palantir Technologies. The system, known as Maven Smart System, fused data from more than 150 sources to generate target recommendations at unprecedented speed. It condensed hours of work into minutes.
But the underlying intelligence on the Minab site was years out of date. Satellite imagery from nearly a decade earlier showed the structure as part of a military compound. In reality, it had long since been converted into a school. The facility maintained an active website with photos of students and posted regularly on social media. None of that information reached the target package fed into Maven.
Bloomberg first reported the details of the unreleased review on September 18, citing multiple officials involved. The findings describe not a single catastrophic mistake but a cascade of preventable failures. Rushed timelines. Reduced staffing for civilian harm mitigation. Outdated records that sat uncorrected in legacy databases. And an overreliance on software that operators mistakenly believed would automatically flag stale or contradictory data.
Some Centcom personnel expected Maven to highlight inconsistencies in the intelligence assembled for potential targets. It is not clear why they held that expectation. The system was designed to integrate and present information for human decision-makers. It was never meant to serve as an independent verifier of source quality. A Palantir spokesperson told Bloomberg the company “is not responsible for the underlying data nor identifying intelligence deficiencies” and said there is no evidence its software was at fault in the Minab strike.
The episode has sharpened debate inside defense circles about the rapid integration of commercial AI into lethal targeting. Maven began as a modest project years ago. Over the past year the Pentagon elevated it to a cornerstone of operations. Several former senior officers with high-level clearances now work for Palantir, some even at Centcom’s Tampa headquarters. The company’s technology helped compress the “kill chain” dramatically during the opening phase of strikes on Iran. More than 1,000 targets were hit in the first 24 hours alone.
Earlier reporting had already pointed to flaws in the intelligence process. A July CNN investigation revealed that senior commanders bypassed explicit warnings embedded in targeting systems about outdated data on Iranian sites. Those warnings required senior officer approval to override. They were overridden anyway. The New York Times reported in March and June that preliminary findings showed U.S. forces were responsible and that the error stemmed from stale records. Iranian state media put the death toll at 168 children and 14 teachers.
The school stood across from a legitimate naval facility. That proximity likely contributed to the initial misclassification. Yet evidence of the site’s true purpose existed in open sources and in at least one disconnected intelligence database as early as 2019. An analyst had noted the change then. The note never migrated to the authoritative targeting database.
Staff cuts compounded the problem. The Trump administration had reduced the number of Pentagon personnel dedicated to assessing civilian risk. With war planning compressed into weeks rather than months, those remaining analysts faced immense pressure. Maven offered speed. Operators embraced it. The system recommended the Minab site as a day-one target. Humans approved it.
Palantir has since updated Maven. The company added new capabilities that re-review underlying intelligence to identify factors that would disqualify a target and flag inconsistencies human reviewers might miss. A person familiar with the changes described them to Bloomberg. Whether those improvements would have prevented the Minab error remains unknown.
The incident ranks among the worst U.S. civilian casualty events in decades. It echoes past tragedies, from the 1991 bombing of a Baghdad shelter that killed more than 400 to repeated strikes in Afghanistan and Iraq where outdated or incomplete intelligence played a role. But the scale of AI involvement marks something new. Lawmakers from both parties have pressed the Pentagon for answers. In March more than 120 Democratic members of Congress sent a letter to Defense Secretary Pete Hegseth demanding details on AI use in target selection and civilian protection measures. They specifically asked whether Maven identified the school and if any human verified its accuracy.
Months passed without full public disclosure. President Trump initially suggested Iran itself was responsible. Later he said it might never be known whose missile caused the deaths amid the chaos of opening salvos. The Pentagon elevated its investigation to a higher-level administrative probe that can support disciplinary action. As of mid-September the final report had not been released.
Defense officials emphasize that AI does not make final targeting decisions. Humans remain in the loop. Yet the Bloomberg account shows how expectations around the technology can distort that loop. Operators began to treat Maven’s outputs as more authoritative than the flawed data beneath them. The system accelerated analysis. It could not replace judgment about what the analysis actually meant.
Critics worry the Minab strike illustrates a broader risk as the military rushes to deploy commercial AI across combat functions. Palantir’s CTO Shyam Sankar told Fox News in March that the Iran campaign demonstrated AI enabling faster planning than in any prior conflict. Strikes per day more than doubled compared with previous operations. That speed came with trade-offs.
Supporters of accelerated adoption argue that human-only processes cannot match the tempo of modern conflict against sophisticated adversaries. Iran possessed advanced air defenses and ballistic missiles. Rapid degradation of those capabilities saved American and allied lives, they contend. The alternative, slower targeting, might have prolonged the war and increased overall casualties.
Yet the human cost at Minab refuses to stay abstract. Photos of the ruined school circulated widely. Mourners buried rows of small bodies. Iranian officials used the deaths to rally domestic support and condemn the U.S. internationally. The episode also fueled online disinformation campaigns, with AI-generated videos falsely depicting the strike in sensational ways, as documented by The New York Times.
Inside the Pentagon the review has prompted quiet reassessment. Officials are examining how intelligence databases talk to one another, or fail to. They are revisiting training on AI tool limitations. And they are weighing whether the pressure for speed in future high-intensity conflicts will inevitably erode safeguards.
Palantir, for its part, continues to expand its defense contracts. The company maintains that responsibility for data quality rests with the government. Two people familiar with its Pentagon agreements told Bloomberg the same. Still, the post-strike software changes suggest the firm recognizes that perception of its tools matters as much as their technical specifications.
The full story of the Minab strike may never satisfy everyone. Investigations remain classified. Some details will likely stay hidden. But the outlines are clear enough. Outdated intelligence. Compressed timelines. Reduced oversight. And a new technology that performed exactly as designed while operators asked it to do more.
That combination proved fatal for 123 children and dozens of adults on a February morning in Minab. The Pentagon now faces the task of ensuring the next AI-assisted targeting cycle learns from this one. Speed without accuracy is not efficiency. It is simply faster error.