A Pentagon Review Reportedly Faults Overreliance on an AI Targeting Tool in the Minab School Strike

An unreleased internal Pentagon review has reportedly found that overreliance on an AI targeting tool was one of the failures behind a US missile strike on an Iranian primary school in February, in which more than 150 people were killed, most of them children.

That finding comes from a Bloomberg investigation published on Friday 18 September, citing officials involved in the review. Bloomberg's page returns 403 to our fetchers, so the Pentagon-review material below is drawn from Gizmodo's report of that investigation and attributed accordingly throughout. The legal findings are quoted from the UN report itself.

One thing to establish first, because the two are easy to conflate. A separate United Nations inquiry has found reasonable grounds to believe the United States committed a war crime in this strike. Its report does not mention artificial intelligence. We checked the document rather than the coverage of it: across its 18 pages, "artificial intelligence", "Maven", "Palantir", "algorithm", "machine learning", "autonomous" and "software" appear zero times, as does "AI" as a standalone word. The two inquiries overlap heavily on the facts — the Mission also finds "the apparent existence of and reliance on outdated targeting data" — and diverge on exactly one thing: whether an AI system is part of the account. Only the Pentagon reporting says it is.

What the review reportedly says about the AI tool

Per Gizmodo's account of Bloomberg's reporting, officials said some personnel inside US Central Command "leaned too hard on the artificial intelligence inside the Maven Smart System, an AI-powered data integration and targeting platform developed by Palantir to accelerate intelligence analysis and decision-making."

The failure described is not a malfunction. It is a mismatch between what the system does and what its users believed it did: officials said "some users expected Maven to flag stale records or contradictions in the intelligence assembled for potential targets, though it is not clear why they thought the system would do that."

The Minab site had been catalogued as an Islamic Revolutionary Guard Corps facility on the strength of outdated data. It went into Maven with other candidates and, in Gizmodo's summary, "came out as a recommended day-one target". Work on target lists that "once took hours was condensed into minutes."

A Palantir spokesperson told Bloomberg 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 the company's Pentagon contracts said the government keeps primary responsibility for the quality of information fed into Maven. After the strike, according to a person familiar with the work, Palantir added features that "re-review underlying intelligence to identify factors that would disqualify a target and flag inconsistencies and inaccuracies that human review may have missed" — functionality said to have already caught some anomalies.

The failures that were not about software

Gizmodo describes the AI issue as one of three elements in the kill chain, alongside bad intelligence and outdated satellite imagery. It does not rank them; in our reading, the other two carry most of the causal weight, and the UN report's own account of the targeting failure supports that.

The school, the UN Mission records, "began operating in approximately 2016, in converted Islamic Revolutionary Guard Corps structures", adjacent to a naval complex whose nearest facility stood about 72 metres away. But the report is emphatic that it had long since been separated from it: "a wall constructed much earlier, sometime between 2013 and 2016, separated the school site entirely from the Islamic Revolutionary Guard Corps complex", and by the time of the strike the school "was operating in a distinct, walled compound with its own entrance and access from the adjacent public street". US databases nonetheless listed the compound as military. Gizmodo likewise reports that construction of walls and separate entrances cutting the school off from the base "appears to have been finished by 2017", and a 2018 image showing painted walls, a football pitch, assembly rows and playground markings; that an analyst noticed changes as early as 2019 and logged them in a system not connected to the primary targeting database, so the notes never reached the people assembling the strike package; and that the school had its own website and appeared on Google Maps.

The Mission notes that this was established early and independently. It records that on 11 March, The New York Times reported a preliminary US military investigation had determined the United States was responsible "after target coordinates were generated using outdated intelligence identifying the school building as part of an adjacent military installation."

Two further factors are organisational rather than technical. More than 1,000 Iranian targets were struck in the first 24 hours of the war, a pace officials said compressed the time available for further confirmation. And staffing on civilian-harm mitigation teams across the Department of Defense had fallen by roughly 90% in recent years, to fewer than 20 people in total; Centcom's team went from ten people to one. No member of those teams reviewed the Minab site before the missiles were launched.

The UN report reaches the same territory by a different route, and puts it more bluntly. It finds the United States "acted wilfully, that is, with the requisite mental element of recklessness (dolus eventualis)", citing "the deprioritization by the Government of the United States of civilian harm mitigation and of compliance with international humanitarian law, expressed through public statements by senior United States leadership, including the President and the Secretary of Defense".

More than 120 House Democrats wrote to Defense Secretary Pete Hegseth in March asking what role Maven and other AI tools played in identifying the site. The Pentagon has not answered publicly, citing the ongoing investigation. Officials told Bloomberg the full report has been substantially complete for months. A Pentagon spokesperson said the incident remains under investigation and declined further comment.

What the UN found, precisely

The Mission's report is A/HRC/63/61, in the advance edited version dated 21 September. It makes two separate war-crime findings about two separate strikes, and they are not the same crime.

On the school, at paragraph 126, the Mission "has reasonable grounds to believe that the strike conducted under the command of the United States on the school building amounts to the war crime of making a civilian object the object of an attack." On a second strike, in Lamerd, at paragraph 128, it finds the war crime of "launching an indiscriminate attack resulting in loss of life or injury to civilians or damage to civilian objects" — that strike hit "a sports centre, an adjacent elementary school and residential buildings", killing 22 people, through the airburst of a Precision Strike Missile dispersing "approximately 180,000 tungsten pellets over a wide area".

The distinction matters, and it has been reported the other way round more than once — including by Gizmodo, our own source for the Pentagon material, which places the indiscriminate-attack finding in the school context — writing, as UN News did, before the advance edited version landed. Earlier summaries, including UN News on 17 September, attached the indiscriminate-attack finding to the school; the advance edited report attaches that to Lamerd and the civilian-object finding to Minab.

There is also a direct line between the two inquiries in this piece. The Mission's finding on the school is based on evidence that "does not include responses by the United States to the Mission's requests for information, or a full and public disclosure by the United States of the outcomes of its internal investigation into the air strike" — on the face of it the same internal review Bloomberg has now reported on.

The Mission made findings against the Iranian government too, and they concern its crackdown on protest rather than the strikes. At paragraph 121 it finds that many gross human rights violations by Iran "amount to the crimes against humanity of murder, imprisonment and other severe deprivation of physical liberty, torture and other inhuman acts committed as part of a widespread and systematic attack directed against a civilian population, namely protesters, including children, their supporters and perceived dissenting or opposing actors …".

The casualty figures, and the school

The report carries two counts rather than one, and does not treat them as a contradiction. At paragraph 92 it records the Minab prosecutor's final toll of 156, including "120 schoolchildren (73 boys and 47 girls), 26 educators, all of whom were women, 7 parents of students". At paragraph 93 it records that "credible independent information asserted that 157 individuals were killed at the school, including 123 children, all aged 13 years or younger, including 15 children who were 6 years old." A Minab Red Crescent member "reported to Iranian media that an additional 110 individuals had been injured".

It also settles a detail that has been reported loosely. This was not a girls' school: paragraph 80 describes "the Shajareh Tayyebeh Primary School for girls and boys aged 6 to 13", and the dead included 73 boys and 47 girls. Al Jazeera reports that in June President Trump said "nobody" purposefully attacked a girls' school in Iran. A senior administration official told Bloomberg that "The United States does not target civilians", and the Defense Department declined to comment on the UN report. Washington has not publicly accepted responsibility.

What this does and does not show about AI in targeting

In our reading, the honest version of this story is narrower than "an AI picked the target", and more troubling than "the software was fine".

Nothing in the reporting says Maven malfunctioned, and Palantir's position is that the data was the government's responsibility. What the officials describe is an expectation gap: users who believed the system would catch stale or contradictory intelligence, when nobody has explained why they believed that. A tool that compresses hours of target-list work into minutes does not thereby acquire the job of auditing its own inputs — but if the people using it assume it has, the human checks that used to occupy those hours stop happening, and nothing takes their place. On the account given, that gap sat alongside a civilian-harm review function cut by roughly 90% and analyst notes stranded in a disconnected database.

That is a procurement and process failure with an AI system at its centre, rather than an AI failure as such. It is also the pattern this desk keeps meeting at smaller scale: we reported this month, following CNN, on an AI chatbot that misread a Chinese-flagged ship's cargo, after which the US military planned to intercept it and then pulled back. The mechanism is the same — a system consulted with more confidence than its design warrants — and the difference is in what was at the other end of it.

Editor's note: Bloomberg published this investigation on Friday 18 September and we are coming to it five days late. We checked our own record before reaching for the obvious excuse of a quiet week, and it is not true. We published nine pieces on 18 September and eight on 19 September. Gizmodo's account — the version we could actually read — went up on Sunday 20 September, a day we published nothing at all. On our reading this is the most significant AI story we have missed this month, and it is a failure of what we chose to look at rather than of capacity.