A Pentagon review names AI overreliance in a strike that killed 123 children
Bloomberg obtained an unreleased internal Pentagon review of the March strike on Shajarah Tayyebeh Elementary School in Minab, southern Iran, on the opening day of the Iran war. Two Tomahawks hit the school. More than 150 people died, at least 123 of them children. The review says Pentagon personnel knew within hours that the strike was American.
It lists a chain of preventable failures: outdated intelligence, compressed targeting procedures, sharp cuts to the staff of civilian-harm specialists, and overreliance on AI. Specifically, some personnel at Central Command leaned on the AI embedded in Palantir's Maven Smart System, expecting the system to flag intelligence that was stale or internally inconsistent. Maven fuses satellite imagery, drone feeds and signals intelligence into strike recommendations at speed, under a Pentagon contract worth up to $1.3 billion.
The failure mode described there is not hallucination and it is not a bad recommendation. It is humans assuming a capability the system was never claimed to have. Nobody sold Maven as a staleness detector. Operators under time pressure, with fewer colleagues whose job was to object, treated speed as validation β if the system produced a recommendation quickly and confidently, someone upstream must have checked the inputs.
That gap between what a system does and what its users believe it does is the thing every deployed agent has, and almost nobody measures. Benchmarks measure output quality. This review is measuring what operators assumed, and the answer was wrong in a way that killed a classroom. The findings land more than six months after the attack, amid congressional pressure for accountability.
https://www.bloomberg.com/graphics/2026-iran-school-attack/
← Back to all articles
It lists a chain of preventable failures: outdated intelligence, compressed targeting procedures, sharp cuts to the staff of civilian-harm specialists, and overreliance on AI. Specifically, some personnel at Central Command leaned on the AI embedded in Palantir's Maven Smart System, expecting the system to flag intelligence that was stale or internally inconsistent. Maven fuses satellite imagery, drone feeds and signals intelligence into strike recommendations at speed, under a Pentagon contract worth up to $1.3 billion.
The failure mode described there is not hallucination and it is not a bad recommendation. It is humans assuming a capability the system was never claimed to have. Nobody sold Maven as a staleness detector. Operators under time pressure, with fewer colleagues whose job was to object, treated speed as validation β if the system produced a recommendation quickly and confidently, someone upstream must have checked the inputs.
That gap between what a system does and what its users believe it does is the thing every deployed agent has, and almost nobody measures. Benchmarks measure output quality. This review is measuring what operators assumed, and the answer was wrong in a way that killed a classroom. The findings land more than six months after the attack, amid congressional pressure for accountability.
https://www.bloomberg.com/graphics/2026-iran-school-attack/
Comments