When the Pentagon did Palantir’s Smart Maven System an official “program of record” in March, it did more than approve the software.
It gave long-term status to a platform that analyzes data from satellites, drones, radars, sensors and intelligence reports to identify potential targets. Maven supported almost all of them 13,500 US attacks in Iran. The designation is intended to secure funding and expand its use throughout the military.
That decision came while the United States was still investigating The February 28 strike at Shajareh Tayyebeh Girls’ School in Minab, Iran. An initial US military investigation found that US forces were likely responsible, but the Pentagon has not issued a final public conclusion.
Iranian authorities have reported between 155 and more than 175 deaths, most of them students and teachers. The question is why a military system designed to move faster from information to action doesn’t have an equally clear public demand to slow down after serious civilian harm.
In software engineering, a “error budget” is intended to protect users. When a service exceeds an agreed failure level, new releases stop until the problem is understood and fixed. War is definitely not software.
But in Washington’s AI-assisted operations, civilian deaths do not automatically trigger a temporary suspension of relevant assets, an independent review of the data, or a public accounting of the decisions that led to the strike.
Proponents of military AI argue that better data and faster analysis can make attacks more precise. This is possible in principle. But accuracy is not something that software simply possesses. It depends on the quality and age of the data, what a system treats as suspicious, and whether people have enough time to challenge its output.
AI does not return stale or biased information to the truth. It can make bad information travel faster.
Human Rights Watch has warned that Israeli digital tools used in Gaza relied on incomplete data and inaccurate assessments that could increase the risk of civilian harm. The lesson is not that every use of AI is illegal. It’s that speed can magnify the consequences of a bad premise before anyone catches it.
The well-known certainty that a “man remains in the loop” does not solve this problem. A person who sees only a final recommendation, cannot inspect the data trail, and has seconds rather than hours to decide, is not exercising meaningful independent judgment.
AI can shape lethal force without pulling the trigger. It can determine what seems urgent, which people or countries rise to the top of a list, and which facts don’t appear at all.
Iran’s campaign makes this an immediate concern. Maven and Claudius of Anthropic are used to process classified intelligence, suggest coordinates, rank targets, and turn planning that once took weeks into real-time operations. This does not prove that the AI independently selected each target or that each AI-assisted strike was illegal.
Reuters reported later that Palantir was faced with the task of removing the Anthropic technology from Maven. But the bigger issue remains: The Pentagon is institutionalizing systems built to compress the time for human deliberation before force is used.
Minab’s tragedy represents the human cost of treating speed as a virtue in itself. School had a public websiteposted pictures of students and was visible in satellite images. He also reported that the targeted officials appeared to have used outdated intelligence.
AI’s exact role in the strike has not been disclosed. But the information apparently did not clearly distinguish the school from the adjacent military complex, and the Pentagon has not publicly explained how the failure traveled through the targeting process. This is a liability problem, not just a technical one.
Congress should enact a civilian ban on AI-assisted targeting. Credible evidence of serious civilian harm should trigger a temporary suspension of the relevant system on comparable missions, an independent review of the data and approval process, and a public explanation of the system’s role.
This would not treat every civilian casualty as evidence of a crime or prohibit any military use of AI. That means families don’t have to wait months in secret to learn whether a database, a contractor or a command decision helped turn a civilian location into a target.
Congress has begun to consider broader guardrails. of House Armed Services Committee Defense Bill of 2027 would direct the Pentagon to expand its AI policy to systems that support or materially affect operational planning, target development, weapons use, and engagement recommendations.
This is a start. But it is not a civil prohibition rule. It does not automatically stop a system after a disruptive incident or warrant a public response.
Such a rule would reinforce the law of war. Customary international humanitarian law requires soldiers to take all possible precautions to verify targets and reduce foreseeable civilian harm. It also requires that they cancel or suspend attacks when it becomes clear that a target is not military or an attack would be illegal.
of Red Cross has recommended after-action reviews and retesting or suspension of AI systems when reliability, security or legal compliance concerns arise. The latest stock market points in the same direction.
Renato Ujku highlights precautions during AI-assisted target verification, while Jessica Dorsey warns that over-reliance on computational systems can undermine the human judgment needed to protect civilians. If an AI-enabled process makes verification more difficult because it rewards speed, the problem is failure to meet an existing legal duty.
Palantir may say it supplies the platform. The Pentagon might say the software only supports decisions. An operator may say that the incoming data is already sorted.
But people mourning outside a destroyed school have no clear answer as to who is responsible. Until Washington establishes a publicly announced civilian ban rule, “just war” will remain a technical phrase masking a political choice: to accelerate force without a meaningful way to stop, inspect, and answer for its failures.
Brian Hudson is a political analyst and freelance journalist.
This article originally appeared on Foreign Policy in Focus and is reprinted under a Creative Commons license
