A US Department of Defense review has concluded that overreliance on artificial intelligence contributed to a March missile strike on a school in Iran that reportedly killed about 120 children, according to the review's findings as reported by Bloomberg on Monday and summarized by Yahoo News and other outlets.
The finding is the first time the Pentagon has formally attributed a deadly combat outcome in part to reliance on an AI system, and it lands in the middle of an intensifying debate over how much targeting authority the US military should delegate to machine judgment. For more reporting like this, see our AI ethics coverage.
What Happened in March
The strike occurred in early March 2026, when video analysis published by Bellingcat and verified by NPR and the BBC showed a US Tomahawk cruise missile striking next to a girls' school in the Iranian city of Lamerd. Within days, The New York Times reported that a preliminary US inquiry had found the strike was the result of a US missile, and Iranian sources put the death toll in the dozens, most of them children.
The new Pentagon review, described in a Bloomberg investigation published Monday, goes further: it reportedly traces the failure up the targeting pipeline and concludes that overreliance on AI contributed to the strike. Yahoo News summarized the review as linking an AI targeting system to the strike that killed 120 Iranian children; several outlets have identified the system in question as Palantir-provided software, though the Pentagon has not publicly confirmed vendor details.
Inside the Kill Chain
The Los Angeles Times, in a parallel investigation published Monday, traced what it called the US "kill chain" that destroyed the school — the sequence of data collection, machine-assisted target nomination and human sign-off that ended in a missile launch. The reporting underscores an uncomfortable reality of modern precision warfare: the crew that fires a Tomahawk may never see the target at all. They execute a target data package assembled upstream, where algorithmic tools increasingly shape what gets nominated and how confident the supporting analysts are.
In other words, the review's central finding — overreliance — is not about a robot pulling a trigger. It is about humans trusting a machine's output enough to skip the skepticism that the process was designed to require.
Oversight That Was Dismantled
Context for the failure was reported months ago. Politico revealed in March that the Pentagon chief had slashed offices not contributing to his "lethality" goals, including the Civilian Protection Center of Excellence, and that staffing on civilian-harm mitigation teams had fallen roughly 90%, shrinking to fewer than 20 people overall. ProPublica reported that Central Command's civilian-casualty team dropped from ten people to one.
Against that backdrop, the review's findings read less like a software bug and more like a systems failure: an AI tool whose errors went unchallenged inside a targeting apparatus whose human safety nets had been cut to the bone.
A Pattern the Pentagon Already Knew About
The Iran strike was not an isolated AI-related near-miss in the US military. Separate reporting by Gizmodo and Futurism documented an incident in which the military nearly boarded — and potentially fired on — a Chinese vessel after an AI system hallucinated nuclear weapons material aboard it, a mistake that could have triggered a great-power escalation. In that case, human intervention prevented disaster. In Iran, no such intervention occurred.
AI researchers and military-ethics scholars have warned for years that machine outputs injected into high-stakes decisions create automation bias — the documented human tendency to over-trust confident systems. The Pentagon's own review now stands as official acknowledgment that this bias contributed to one of the war's deadliest civilian tolls.
What Comes Next
The review is likely to intensify pressure in Washington for statutory human-in-the-loop requirements in AI-enabled targeting, at the same time as the administration pushes to accelerate military AI adoption. It also complicates the sales pitch of defense-tech vendors whose products are now implicated in a formal fault review — and it gives Congress a concrete, documented case study to cite in hearings on autonomous weapons policy.
Many questions remain unanswered. The review's summaries published so far do not specify which model version was in use, whether the failure was a hallucination, a misclassification of the building, or corrupted input data upstream — distinctions that matter enormously for how the military should regulate AI tools. Nor is it clear whether the system's outputs were reviewed by any analyst before the target package was finalized.
Public opinion is already skeptical: a Reuters/Ipsos poll published this week found 73% of Americans say AI firms are not doing enough to prevent catastrophic outcomes, and concern about AI in weapons systems consistently polls higher than concern about consumer AI. What happens to the review's recommendations — and whether anyone is held accountable — will signal whether the finding changes practice or becomes another incident report filed and forgotten.
---
Get the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
