What the investigation found
A Bloomberg investigation published on Friday reports that over-reliance on Palantir’s Maven Smart System was among the causes of the 28 February strike on the Shajareh Tayyebeh elementary school in Minab, in Iran’s Hormozgan province. More than 150 people were killed, at least 123 of them children, Bloomberg reported.
The finding is not that the software chose the target on its own. It is that the people using it believed it was doing work it was not built to do. Personnel “over-relied on and over-trusted” Maven, according to ThePrint’s account of the investigation, expecting the system to flag outdated information and internal inconsistencies that it was never trained to detect.
A cascade, not a single failure
Investigators described a chain of preventable errors rather than one bad output. The satellite imagery used to assess the site had not been updated for seven years. The site’s classification was years out of date. Decision timelines had been compressed by an order for a large-scale aerial campaign.
And the human check that might have caught the rest had been thinned out. Bloomberg reports that staffing on the Pentagon’s civilian harm mitigation teams had been cut by roughly 90%, and that nobody from those teams reviewed the strike plan before it was carried out.
Why speed is the mechanism
The detail that explains the rest is timing. Maven condenses an assessment that traditionally ran through several stages over hours into a process that takes minutes. That is the reason militaries buy it.
It is also the reason a stale classification can survive all the way to a launch decision. When a process is fast, the steps that used to create friction — and with it, the chance for someone to notice that a satellite picture is seven years old — are the steps that get removed.
US officials cited by Bloomberg described overconfidence in AI tools and a misplaced belief that those tools had capabilities beyond their intended scope. That is a description of a training and doctrine failure as much as a software one.
What Palantir says
Palantir told Bloomberg there is no evidence its software was at fault, as summarised by AI Weekly. The company has since added disqualifier flags to Maven — a change that, read plainly, adds a way for the system to mark information that should stop a target from advancing.
Both things can be true. A tool can perform exactly as specified and still contribute to a catastrophe, if the people around it have been told, or have assumed, that it is checking something it is not.
What to watch
The practical question now is whether the disqualifier flags are a product change or a doctrine change. A flag only helps if the workflow stops when it fires, and if the staff who would act on it still exist. Watch for whether the Pentagon restores the civilian harm mitigation headcount, and whether any published standard requires an AI-assisted targeting system to state the age of the data it is reasoning over.
