On March 3, a mass funeral was held for over 150 students and staff killed in the February 28 US missile strikes on the Shajareh Tayyebeh girls’ elementary school in Minab, Iran.IRCS/Zuma
More than 60 percent of Americans believe it’s at least moderately likely that artificial intelligence will kill us all, according to a Politico survey released Wednesday.
It’s easy to see why. Arguments that AI companies are “gambling with our lives” have dominated the news this week. But despite our fear that AI might kill us en masse, another new poll shows that very few Americans know our government is already using AI to do just that abroad.
The US-Israel war on Iran is not only the least popular war the United States has ever waged, but also represents an unprecedented application of AI in warfare. New data from the Institute for Global Affairs and YouGov shows that the present weaponization of AI has flown under the radar, with most Americans largely unaware of our own military’s AI use.
“People would be pretty shocked that what they’re using for their homework or to summarize emails is also being used by intelligence officers.”
Only 8 percent of Americans have heard “a lot” about the military’s use of AI in the war with Iran. The majority, almost two thirds, have heard “only a little,” and four in ten have heard “nothing at all.”
“We wanted to get a gut check on, do Americans know this is happening?” survey co-author Jonathan Guyer said in an interview. Despite a long history of US military development eventually boomeranging back home, the answer to that question might be “no”—notwithstanding Defense Secretary Pete Hegseth’s relentless messaging about the “AI-first Department of War.”
“People would be pretty shocked that what they’re using for their homework or to summarize emails is also being used by intelligence officers,” Guyer said.
Cameron Stanley, the chief digital and artificial intelligence officer for the Department of Defense, wrote in June that xAI’s Grok “enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours during Operation Epic Fury.”
That admission, part of the administration’s case for the national security importance of Elon Musk’s data centers, is just the tip of the iceberg. The Pentagon’s Maven Smart System, built by Palantir, also uses Anthropic’s Claude AI—the same program that people use to create AI boyfriends for themselves or draft awkward texts to their roommates—to “semi-autonomously rank targets by strategic importance, drafting automated legal justifications for each strike along the way,” according to Katie Livingstone of Military Times.
Claude’s target selection is reportedly only about 60 percent accurate in good conditions, 30 percent accurate in bad weather. So Pentagon officials claim there’s always a human in the loop when selecting who to bomb.
“We follow the law and humans make decisions,” Hegseth told the Senate Armed Services Committee in May. “AI is not making lethal decisions.”
But it’s making lethal recommendations, which don’t appear to receive enough human oversight to prevent atrocities. In February, the US military fired Tomahawk missiles directly at a girls’ school in Minab, Iran and killed over 150 people—most of them young children. Seven months later, UN experts called the strike a crime against humanity.
As AI companies enter the battlefield, “the scale of atrocities committed…is rising correspondingly.”
In the immediate aftermath of the attack, much of the US news coverage focused on whether Claude was used to select the school as a target. But the fact that a human was at some point involved does not lessen the harm done.
At a congressional human rights hearing September 16, the Carnegie Endowment for International Peace’s Steven Feldstein explained how AI has facilitated a massive ramp-up in the speed of missile target selection.
“In the Iran war, during the first week of hostilities alone, the Pentagon reported it had struck more than 3,000 targets,” Feldstein said. By the time a ceasefire was announced 38 days later, that number of targets had risen to 13,000.
“In Israel, AI targeting systems have played a crucial role in supporting its military campaign against Gaza,” Feldstein added. One system, known as Lavender, identified 37,000 individuals as potential airstrike targets in the initial weeks of the war. As AI companies make their way onto the battlefield, Feldstein said, “the scale of atrocities committed by warring parties is rising correspondingly.”
In January of this year, Pete Hegseth released a memo directing the United States to embed AI in every part of the “kill chain.” As it moves to comply, the Department of Defense is placing relatively few checks on what technology companies can or cannot do when the United States is waging war. “We’re in a situation where the companies are essentially regulating themselves,” Guyer, of the Institute for Global Affairs, said.
Ethical principles for AI use that the Department of Defense adopted in 2020 require only that military AI be “governable” by humans. And the additional safeguards AI policy groups tend to propose revolve around stopping fully automated technology—not regulating the semi-autonomous technology that is already being used to kill at scale.
As American policymakers debate potential future AI threats, it seems as though the immediate threat of AI—the ongoing expansion of those “kill chains”—is not even making it into the AI safety discussion. Nor are the victims of AI-assisted warmaking.
That means tech companies are still the ones making the rules: “Their terms of service are defining how AI systems are being used for war to a greater extent than any laws that have been written, or to a greater extent than even executive branch guidelines,” Guyer said.
