The Pentagon says Grok helped strike 2,000 targets in 96 hours, and that the polluting power plant behind it is a matter of paramount national security. The two claims belong in the same sentence, and that is the problem.
The admission did not come in a press release or a Pentagon briefing, it came in a statement filed in a federal courthouse in Mississippi, in a case about air pollution. There, defending Elon Musk's xAI against a Clean Air Act lawsuit, the Pentagon's chief digital and artificial intelligence officer set down a sentence that ought to stop anyone reading it cold.
The chatbot known as Grok, he wrote, had helped fire more than 2,000 munitions at 2,000 distinct targets in Iran within 96 hours, and its continued operation was a matter of paramount national security.
Read that again, because the venue is the whole story. A government official disclosed that a consumer AI product had been used to bomb a country, not in order to inform the public, but in order to keep a data centre running. The targeting and the turbines arrived in the same affidavit because, in the administration's telling, they are the same argument: Grok matters to the war, the war matters to the nation, therefore the power plant that feeds Grok cannot be turned off, whatever the law says about its permits.
This is the moment the two problems with military AI stop being separate. The first is that a chatbot built to answer questions on a social network is now wired into the machinery that kills people. The second is that the man who owns that chatbot also sits inside the government deciding how it gets used. Each is alarming on its own.
Together they describe a system in which the safeguards we would normally expect, legal, environmental, ethical, all bend toward the convenience of one company and one man.
Start with the chatbot. Grok is, according to the filing, one of only four AI models the Pentagon considers capable of supporting national-security applications, and one of three cleared for mission-critical work in top-secret settings.
It feeds into the National Geospatial-Intelligence Agency's Maven Smart System, the AI-driven dashboard that lays out intelligence data to help officials decide what to strike. The official line is reassuring in its phrasing: the AI does not create targets, it identifies points of interest for human analysts to weigh. The humans, we are told, remain in the loop.
The loop did not save the children of Minab. On 28 February, in the first wave of strikes, a Tomahawk cruise missile hit the Shajarah Tayyebeh girls' school in the Iranian town, killing scores of people, more than a hundred of them children by most counts, in what observers have called the deadliest single civilian incident of the conflict.
A preliminary US military inquiry has concluded that American forces were likely responsible, and that the strike rested on outdated intelligence: target coordinates built from stale data, fed through the system, never rechecked against a current map. The school had been a school the whole time.
That is the trouble with putting AI near targeting, and it has nothing to do with whether a machine pulls the trigger. The danger is subtler and harder to legislate against. An AI dashboard that ranks points of interest and pairs weapons to them does not remove human judgement so much as launder it.
The analyst sees a confident interface, a clean readout, a target the system has already dignified with a number, and the natural human response is to trust it. Automation bias is not a bug in these systems; it is the predictable result of building something designed to look authoritative. When the underlying data is wrong, the interface makes the error look like rigour right up until the missile lands.
Now add the second problem, the one the Mississippi filing makes impossible to ignore. The lawsuit was brought by the NAACP, which alleges that xAI is running dozens of gas-burning turbines to power its data centres without the permits the Clean Air Act requires, in a case centred on a Southaven, Mississippi facility where the group counts 27 unpermitted turbines sitting beside homes, schools and churches in a largely Black community.
Memphis, where xAI's Colossus supercomputer hums, ranked second in the country for asthma-related emergency-room visits in 2024. These are the people breathing the exhaust of the compute.
And the government's response to their lawsuit was not to defend the permits, which do not exist, but to argue that the pollution is vital to war. The Department of Justice intervened to ask a judge to throw the case out, on the grounds that shutting the turbines would “severely” impair the Pentagon.
The data centres, the official wrote, are well positioned to provide a critical surge in energy capacity in the event of armed conflict. National security, in other words, has become the universal solvent: it dissolves environmental law, it dissolves the objections of the families downwind, and it does so in service of a private company's unpermitted infrastructure.
Here is where the conflict of interest stops being abstract. Musk is not a vendor at arm's length from the state. He has occupied a singular position inside this administration, and his companies have woven themselves through its security apparatus: xAI's Grok on classified networks, SpaceX's satellites carrying the data, the same man's interests on both the selling and the buying side of the deal.
When the government argues that Grok is indispensable and its power plant untouchable, it is arguing for the commercial interests of an insider. The national-security claim and the business interest point in exactly the same direction, every time, and there is no one in the room whose job is to notice.
The contrast with how this administration treated a more cautious company could not be sharper. When Anthropic refused to let its Claude model be used for domestic surveillance or fully autonomous weapons, the Pentagon designated it a supply-chain risk, a label normally reserved for foreign adversaries, and the company is now locked in litigation over it.
The lesson the market just learned is unmistakable. Insist on guardrails and you are a national-security threat. Run unpermitted turbines and let your chatbot help bomb a school, and you are a national-security asset. The safety-conscious firm gets blacklisted; the compliant one gets its pollution reclassified as patriotism.
Some in Congress have seen the shape of this. Senator Kirsten Gillibrand has introduced the Secure and Accountable Military AI Act, which would keep human decision-makers in control of high-consequence calls and bar AI outright from nuclear, domestic-surveillance and autonomous-weapons uses.
“The most critical decisions affecting our national security and the lives of our service members must always be made by human beings, not unaccountable machines,” she said.
It is the right instinct. It is also, on the evidence of Minab, an incomplete one, because the danger was never only the fully autonomous machine. It was the human who trusts the confident screen, and the company that profits whether the screen is right or wrong.
We have arrived somewhere genuinely new, and we arrived by accident, through a pollution lawsuit. The same product that answers questions about football fixtures and posts jokes on a social network is now rated among the handful of systems trusted to help direct a bombing campaign.
The same turbines that are poisoning the air in South Memphis are, by the government's own argument, an instrument of war. And the same man owns all of it, while helping to govern the state that buys it. None of this was decided in the open. It surfaced because a civil-rights group asked a court to make a company obey the Clean Air Act, and the government answered that it could not, because the chatbot had a war to fight.
The children of Minab cannot be brought back by better software or a stricter statute. But the question their deaths force is still in front of us, and it is not really a question about machines. It is a question about who we let stand between a confident interface and a missile, and whether we are willing to call a thing dangerous while its owner is still in the room telling us it is indispensable.