Pentagon Investigation Finds AI Overreliance Contributed to Strike Killing 123 Children at Iranian School
A Pentagon investigation into a missile strike on a girls’ school in Minab, Iran has found that overreliance on AI targeting systems was among the contributing factors in one of the deadliest targeting errors of the US-Iran conflict. 123 children were killed. Bloomberg published the investigation on September 18, 2026, based on accounts from officials involved in the unreleased internal review.
The investigation identified three compounding failures: faulty intelligence about the target, imagery that had not been updated before the strike was authorised, and AI systems that were trusted over human review at multiple points in the decision chain — a process the US military calls the “kill chain.”
What the Investigation Found
Pentagon investigators concluded that the strike proceeded under time pressure that compressed the normal targeting review process. AI-generated targeting assessments were passed up the chain without sufficient human verification at key checkpoints. When the imagery used to identify the target was examined after the strike, investigators found it predated the school’s construction on the site.
The details emerged through officials who spoke to Bloomberg on condition of anonymity from an internal review that the Pentagon has not publicly released. Palantir, whose Maven Smart System was used in the targeting process, issued a statement defending the software. Previous incidents — including the near-boarding of a Chinese vessel after an AI system hallucinated a nuclear threat — involved errors caught before casualties occurred.
Implications for AI in Military Operations
The strike occurred at the start of the US-Iran conflict. The Minab investigation is among the highest-profile incidents in a growing body of documented cases where AI-assisted military decision-making has produced outcomes that independent human review would likely have flagged.
The Pentagon has not publicly released the formal investigation report. Palantir — whose Maven Smart System powered the targeting process — and Anthropic, whose Claude model is integrated into Maven, were both named by Bloomberg. Palantir issued a statement to Bloomberg defending the software and arguing it bears no responsibility for underlying intelligence deficiencies.
The incident is likely to accelerate legislative pressure around AI use in military targeting. A White House framework for commercial frontier AI, announced in August, called for a voluntary review window of up to 30 days for covered frontier models — but it applies to commercial AI deployment, not military targeting systems.
Context
Earlier this year, Anthropic CEO Dario Amodei stated publicly that the company did not know exactly how Claude had been used in connection with the same strike. That statement came in June; the Bloomberg investigation published September 18 provides the first detailed account of what Pentagon investigators actually found.
The Minab school strike is one of the worst single incidents in terms of civilian casualties attributed in part to AI-assisted targeting. The combination of outdated training data, compressed decision timelines, and insufficient human verification checkpoints echoes concerns that AI safety researchers have raised about high-stakes autonomous systems for years.