A sworn declaration says xAI's Grok sat inside the military's kill chain, the sequence of decisions that turns a sensor reading into a strike order, and helped direct 2,000 munitions at 2,000 distinct targets in Iran over 96 hours.
The Pentagon's chief digital and artificial intelligence officer said in a sworn court filing this month that xAI's consumer chatbot Grok was used to direct roughly 2,000 munitions at 2,000 distinct targets in Iran during a 96-hour window of the war.
The disclosure came from Cameron Stanley, the official who runs the Pentagon's AI portfolio, in a declaration filed in Mississippi federal court defending xAI against an NAACP Clean Air Act lawsuit over a 57-turbine data center being built in a Black community Cameron Stanley sworn declaration, Mississippi NAACP v. xAI. The Independent was first to report the filing's contents, and Gizmodo's read of the same declaration traced it into public view The Independent Gizmodo.
Read literally, the figure describes a military kill chain (the sequence of decisions and data flows that turns a sensor reading into an order to fire) operating at a tempo and scale that no human team could match. The 108-day war, Operation Epic Fury, opened on 28 February 2026 and ended with a memorandum signed in Pakistan on 13 June. The Grok window falls inside it. The deployment has not been confirmed through a separate Defense Department readout or a CENTCOM statement, and the same filing supplies the same "paramount national security" framing the War Department has used in other venues, so the operational claim should be read as the government's own characterization rather than independent confirmation.
What sits behind that claim matters more than the number. The targeting platform the U.S. military used in this war is not a chatbot. It is Maven Smart System, Palantir's software, run under the National Geospatial-Intelligence Agency, that fuses satellite imagery, drone feeds, signals intelligence, and human reports into strike packages The Guardian, March 2026. The Guardian's long read on the Minab school strike argued that the public narrative of "Grok for Iran, Claude for Minab" misattributes the kill chain. A DIA database had classified the Shajarah Tayyebeh primary school as an IRGC-adjacent military facility. Nobody had updated the record since 2016, when the school was built on the same site. The school was misclassified; the strike was a Maven output. Whether Grok's role was target nomination, prioritization, or battle-damage assessment is the open question, and the filing does not specify.
In January 2026, Defense Secretary Pete Hegseth stood up what the War Department now calls an "AI-first" warfighting doctrine, which is shorthand for treating artificial intelligence as a primary instrument of combat rather than a support tool War Department, AI Acceleration Strategy release, January 2026. Grok is one of two frontier models in the DoW's GenAI.mil enterprise program, alongside Google Gemini, deployable at Impact Level 5 and above, meaning it can be used for unclassified through controlled unclassified information. The chain that runs from a commercial chatbot API to a strike order passes through a procurement process that was not designed to assign blame when a target turns out to be a primary school.
That is the real story the filing exposes, and it is the one Stanley's declaration does not address. When a consumer chatbot sits inside the targeting loop, accountability dissolves across at least four layers: xAI, the company that built and continues to update the model; Palantir and the National Geospatial-Intelligence Agency, the integrators who decide what the model sees; the War Department's chain of command, which sets policy and approves targeting authorities; and the individual warfighter who signs off on each strike. Each layer can point to the others. The chatbot has no mechanism to refuse a target and no name to put on a strike record. The War Department will not say who in this chain approved the 2,000 targets, who audited the strikes, or whose name appears on the kill chain for any one of them.
The legal record around military AI is starting to catch up. Senator Kirsten Gillibrand has proposed legislation that would require human control of life-and-death decisions and bar AI use in nuclear command, domestic surveillance, and fully autonomous weapons. Anthropic's Claude is now at the center of a separate legal fight: the DoD designated Anthropic a "supply-chain risk to national security" after contract talks collapsed over Claude's use in autonomous systems and domestic surveillance, and the company is challenging the designation in court. None of those venues has the facts of this filing in front of it yet, and the accountability gap the filing describes is the gap those proposals and that litigation are trying to close.
The harder question is not whether a chatbot should be allowed inside a kill chain. It is what "accountability" means in a chain where the model is a consumer product, the targeting record is classified, and the integrator is a private contractor. Stanley's filing answers a 96-hour question with a 96-hour number. The 175 to 180 people, mostly girls aged 7 to 12, killed at the Minab school on 28 February deserve an answer that lasts longer than that.