The deadliest American targeting error of the century now has an institutional answer, and it names the AI. Pentagon investigators — briefed to officials and still unreleased — concluded that overreliance on Palantir’s Maven Smart System, alongside stale intelligence, outdated imagery, and a civilian-harm-mitigation team cut from ten people to one, sent a school in Minab onto the strike list last February. Two Tomahawks hit it; more than 150 people died, at least 123 of them children. Seven months later, CENTCOM has said it is rewriting how AI selects and vets lethal targets — adding building-classification checks to Maven and live civilian-presence tracking. The probe report stays classified; the process change does not.
The tool that couldn’t flag its own stale data
What happened. Bloomberg’s reconstruction, confirmed by officials with direct involvement in the internal probe, puts Maven — Palantir’s AI command-and-targeting platform that fuses more than 150 data inputs — at the center of the Minab kill chain. The site’s classification as an IRGC facility was years stale; visible school construction dated back around a decade, and one analyst logged the changes in 2019 into a system investigators say wasn’t connected to the primary targeting database. In the compressed hours of the February offensive, Maven surfaced Minab as a recommended day-one target, and “high confidence” held through every review stage. The UN’s Independent International Fact-Finding Mission on Iran concluded the US “failed in its obligation to do everything feasible to verify” the school was a military target — a failure it said “went beyond negligence.” Palantir disputes fault, saying it is “not responsible for the underlying data,” and the Pentagon said the incident remains under investigation. Separately, CENTCOM said it has implemented upgrades to Maven including real-time classification of building functions and civilian presence near potential targets.
Why it matters. The AI-in-the-loop question in targeting has been theoretical for a decade; this is the documented version, at the top of the severity scale, with the vendor named by its own customer’s investigation. The failure is not a hallucination in prose — it’s automation bias against a stale dataset with no breaker: analysts and commanders stayed “high confidence” because the tool appeared to have done the review, when the tool’s job was to accelerate, not to doubt. Anything that puts an un-doubted feed in front of confident humans inherits this failure mode. The institutional fix is also the tell: the Pentagon is patching Maven with the checks a human chain was supposed to perform, which is the same shape as every safety retrofit in engineering — the loop gets rewritten after the body count, not before.
What I’m watching
Whether Palantir’s no-fault position survives the contract-and-procurement cycle rather than the finding — the report stays classified, but Maven’s public upgrade list is an acceptance of scope. And the thread is continuous: from the hallucinated manifest to Minab, when AI accelerates a decision loop without a mechanism to doubt its own inputs, the false positive eventually stops being a ship-manifest near-miss and reaches a target list. The week’s other armed thread — the closed labs’ pace-and-price fight — got its answer in one Tuesday of token cuts; the military’s version just shipped as a process change.
Source: bloomberg.com, gizmodo.com, cryptobriefing.com