UN Finds 'Reasonable Grounds' the Minab School Strike Was a War Crime — and the Kill Chain Was Made in Silicon Valley

Seven months after two Tomahawk missiles collapsed the roof of the Shajareh Tayyebeh primary school in Minab, Iran, the question of legal accountability has finally been answered — even as the Pentagon’s own investigation remains stuck in limbo. On September 17, an independent United Nations fact-finding mission concluded there are “reasonable grounds to believe the United States committed the war crime of launching indiscriminate attacks resulting in the loss of life or injury to civilians or damage to civilian objects.” Bloomberg published its own forensic reconstruction of the kill chain the next day. Together, the two documents mark the most consequential moment yet in the debate over AI-assisted military targeting — because both point, from different directions, at the same machinery.

What the UN found

The mission, which will submit its report to the UN Human Rights Council in Geneva, examined US conduct during the opening hours of the 2026 Iran war. Its findings on the February 28 school strike are unambiguous in structure:

• The school was “clearly identifiable.” UN experts said they found no evidence the building was being used for military purposes. Satellite imagery from December 2025 — two months before the strike — showed dozens of people, apparently playing, in the school’s courtyard.
• The US relied on stale intelligence suggesting a senior Iranian commander was on site and failed to adequately verify it before launching the strike. Imagery from 2013 did show the school co-located with an IRGC naval base; but images from 2016 already revealed a fence separating the school from the base and a separate school entrance. “The US failed in its obligation to do everything feasible to verify that it was a military objective,” the report found.
• The failure was “more than mere negligence.” The mission concluded the US “directed the strikes at the building of the school while being aware of a substantial risk of striking a civilian object and acting recklessly as regards the possibility that this would happen.”
• A second strike the same day on a sports complex and residential area in Lamerd, using a weapon that scattered tungsten pellets, killed 22 civilians and was likewise found to be an indiscriminate attack amounting to a war crime.

At least 156 people were killed at the school, including 120 children and 26 teachers, according to Iranian officials. The White House rejected the findings, and a State Department official told CNN the administration does “not give credibility” to the report — noting the US formally withdrew from the UN Human Rights Council in February 2025. The report could nonetheless be used by international courts including the ICC, and the US maintains its own war-crimes statutes.

The kill chain, reconstructed

The UN report establishes the legal character of the strike. Bloomberg’s investigation, published September 18, reconstructs how it actually happened — and why “AI error” is simultaneously too simple and too generous an explanation.

The mechanical facts were known by March. US Central Command created the strike coordinates using outdated information provided by the Defense Intelligence Agency. The school had been partitioned from the adjacent military base in 2016, yet the site remained in military databases as an active target. The result was a “targeting fiasco,” in the words of one preliminary-probe description, executed with picture-perfect precision against a building full of children.

What Bloomberg’s reconstruction adds is the system-level view. The war’s targeting workflow ran through Palantir’s Maven Smart System — the Pentagon’s flagship AI targeting platform, which uses Anthropic’s Claude to semi-autonomously rank candidate targets. CENTCOM has confirmed US forces used “a variety of advanced AI tools” during the war. The speed of that pipeline, analysts told Bloomberg, was itself a factor: overreliance on machine-ranked recommendations, layered on flawed and outdated underlying data, eroded the human verification that international law demands. In June, Anthropic CEO Dario Amodei told Bloomberg he did not know what role Claude played in the strike.

That last detail matters for the AI industry beyond this war. The blog has previously covered how AI-assisted target generation is scaling errors at machine tempo — 20 soldiers now doing work that once took 2,000 analysts, with programs pushing toward 1,000 tactical decisions per hour. Minab is what that tempo looks like when the data layer is rotten. The model didn’t need to misidentify anything; it inherited a database that said “military target,” ranked accordingly, and the humans riding the kill chain at machine speed had no realistic chance — or took none — to catch the 2016 fence.

Accountability in limbo

The UN’s judgment lands on a Pentagon process that has stalled. CNN reports that the first two stages of the standard three-stage investigation were completed within a week of the strike — confirming US responsibility and that the intended target was struck. But the third stage, in which DIA analysts review satellite imagery and intelligence sources to produce a holistic determination, has still not been ordered. A separate investigation including interviews with service members involved was launched, but its findings have been “locked down” by CENTCOM, with only a handful of officers permitted access. Senior US commanders had been warned before the war that targeting data was severely out of date and approved strikes anyway, CNN previously reported.

The White House response followed the pattern. President Trump initially suggested Iran might be to blame for the strike, then conceded in June that “nobody” had purposely attacked a girls’ school, citing an investigation whose preliminary findings have never been published. CENTCOM said in March that US forces had not struck Lamerd at all on February 28 — a claim the UN report now contradicts.

Why this is the industry’s problem

The UN report is about the United States military. But the accountability gap it exposes runs directly through Silicon Valley, and the industry’s current posture suggests it is not prepared for what comes next.

Anthropic’s Claude sits inside Maven as the ranking engine for target candidates. When Amodei says he doesn’t know whether his model was used in a specific strike, that is not just a public-relations liability — it is an audit failure. A lab that cannot observe how its model is used inside a military system cannot verify compliance with its own usage policies, cannot cooperate with an investigation, and cannot improve the system. The same structural opacity now applies to every frontier lab selling into defense: OpenAI’s expanding national-security stack, Meta’s open weights repurposed by third parties, Google’s Defense Polestar work. Palantir, for its part, told Bloomberg there is no evidence its software was at fault and has since added “disqualifier flags” to Maven — a tacit acknowledgment that the system’s screening needed exactly the kind of civilian-status check whose absence killed 123 children, per Bloomberg’s count, or 120, per the UN’s.

There is also the uncomfortable timing: the industry is, this same week, publicly debating whether to slow frontier development, with Anthropic’s own CEO among those calling to pace the frontier — even as Reuters reports the company weighs releasing a new model before its IPO. The Minab findings are a reminder that the most acute AI risks of 2026 are not hypothetical future superintelligence. They are present-tense, well-documented, and measured in graves: a system that accelerated decisions past the speed of verification, deployed on unverified data, with no functioning accountability loop after the fact. The Sanders-Casar superintelligence-ban bill polling at 68% support exists in this climate for a reason.

What to watch

• Whether the UN Human Rights Council formally transmits the report, and whether any jurisdiction attempts to act on it.
• Whether the Pentagon ever orders the third stage of its investigation — the one that would explain, rather than merely confirm, the strike.
• Whether US labs adopt contractual or technical audit rights for military deployments of their models, as opposed to disclaiming knowledge after the fact.
• Whether “meaningful human control” — the standard invoked by both the UN report and this week’s US-China AI-safeguard proposals — acquires an operational definition that includes data-currency requirements, not just a human signature on the firing decision.

The school was blue and pink, with sports fields painted on the asphalt. Everything downstream of the stale database performed exactly as designed. That is the finding worth keeping: the war crime, in the UN’s framing, was not a malfunction. It was the system working as built.