The US government has empowered America’s AI giants to set their own rules for their models’ use in warfare, without asking or informing the public.

Claude has helped select and prioritize targets in Iran. Grok coordinated the thousands of missiles it launched in the first days of the war. And ChatGPT is a veteran of Israel’s Gaza war.

There’s something absurd about the fact that AI models that most users have tasked with writing grocery lists or summarizing emails are undergirding the new “forever wars” in the Middle East. Yes, Grok, it’s true: most Americans—who live in a country with one of the highest penetration rates of AI use anywhere in the world—don’t realize that these consumer technologies are also powering President Donald Trump’s ongoing war in Iran. AI-enabled technologies have supercharged American military operations, yet 42 percent of Americans have heard nothing about it.

AI Companies—Not the Public—Are Setting Their Own Warfare Rules

My organization, the Institute for Global Affairs, worked with YouGov to field a nationally representative survey of voting-age Americans on the Iran war. Only 8 percent of Americans have heard a lot about AI’s role in the war, while a majority of Americans (63 percent) have heard “only a little” or “nothing at all.”

Though many of these military tech companies advertise their advanced algorithmic warfare on Washington, DC, buses or at Reagan National Airport, the public has been excluded from the conversation. Last year, I argued in the Financial Times that the public deserved answers; our own polling showed that a majority of Americans want stricter rules around AI’s military uses.

Though The Washington Post and other news outlets have since reported on how extensively AI was being used in Trump’s wars, and Pentagon leaders have confirmed that those systems are “active,” those findings have not broken through with voters. The knowledge hasn’t produced guardrails, and the ones that exist on paper aren’t being implemented, according to an inspector general. In fact, what actually has set the standards has been the terms of service of the companies in question.

That AI companies are, in effect, governing their own uses—rather than lawmakers or the executive branch—presents a new, perhaps even more troubling reality. 

This came to the fore when an Anthropic executive asked whether its LLM, Claude, played a role in the Venezuela raid in which Nicolás Maduro was kidnapped. One month later, Anthropic refused the Pentagon’s demand for “all lawful uses” of the chatbot, leading to it being designated as a supply chain risk. (A judge overturned the designation in August after Anthropic sued over the contract.) In contrast, xAI, the maker of Grok, signed on to the “all lawful uses” clause. The Pentagon continued to run Claude through the interface it uses called the Maven Smart System, according to the data-processing mega-company Palantir, which runs Maven.

On the other side of the war, Iran is using the same technology. Iranian actors, according to Anthropic’s recent threat assessment memo, tried to target US bases with an assist from Claude, though Anthropic blocked this use. Tehran is all but certain to try again.

OpenAI, for its part, quietly deleted its own ban on “military and warfare” from its terms in 2024, as The Intercept’s Sam Biddle uncovered. More recently, the Pentagon asked OpenAI to create a chatbot with “minimal refusal rates.” (Both OpenAI and the Defense Department deny this and say it was draft language.)

The Steep Human Cost of AI Errors

Does the tech work or not, or does it work too well? Both, and both are scary. Each of these consumer chatbots likes to remind us that it can make mistakes. As of 2024, Maven’s accuracy is about 60 percent overall, per the Army officer running it—or about 24 percentage points lower than a human.

Military targeting mistakes, whether by a human or a machine, have tragic real-world consequences. In the first hours of the US and Israel’s war on Iran, a missile strike on a girls’ school in Minab, Iran, killed 175 people, overwhelmingly schoolchildren. It was the deadliest incident of the war to date, and yet only around half of Americans have heard about it, according to IGA polling. In March, 120 House Democrats wrote a letter to the Pentagon asking about whether AI, including Maven, was used in the strike.

The significant shortfalls of such technologies have been on display throughout Israel’s war in Gaza, which some UN experts and Israeli human rights groups have described as a genocide. Israel called its May 2021 Gaza operation “the first AI war.” Now, the extent of AI-powered targeting is at a much more drastic scale, with +972 Magazine reporting that Israel’s Lavender and Gospel systems have created tens of thousands of targets—and in turn loosening the tolerance for civilian deaths. Israeli military personnel told the scholar Sophia Goodfriend that ChatGPT and other tools have been used in Gaza. Still, nearly three years after October 7, it is hard to argue that the AI war has worked very effectively for Israel, leading to mission creep and further destruction in Gaza.

America’s Drone Wars Foreshadowed the AI Wars

There is an American precedent for this, too. The “drone wars,” first launched under George W. Bush and accelerated under Barack Obama, offered a tantalizing vision of unmanned operations launched from afar. Congress never legislated on drone strikes, and the executive constraints that Obama’s team created were insufficient; there was little incentive to regulate unmanned warfare. Nor was there any push through international arms control or the UN.

The latest generation of drones, many with autonomous capabilities, are the weapons of Trump’s wars. This month, Iran reportedly captured an underwater drone from US unmanned systems developer Anduril. An advanced drone from Shield AI, another major US military tech startup, was spotted as part of the CIA’s boat strike campaign in the Caribbean, Drop Site News reported.

Throughout this period, the only backstops have been the billionaires who founded the companies, the tech titans on their boards, and maybe the Gulf money with government links that funds them through investment rounds. The American public has never received anything resembling transparency for these efforts—whether that be mandatory disclosure to Congress when AI is in the kill chain, a human-review requirement for launching strikes, or the full findings of the Minab school bombing investigation released to the public. In 2026, you’re more likely to hear an on-the-record briefing on AI and drone effectiveness on a quarterly earnings call than in a Senate committee.

This problem extends well beyond unproven technology. When the check on AI’s power to kill comes from a private company’s terms of service, it makes it far easier to slide into war.

About the Author: Jonathan Guyer

Jonathan Guyer directs the Independent America program at the Institute for Global Affairs at Eurasia Group, where he runs national surveys on war, military spending, and America’s role in the world. He hosts the podcast None of the Above. A contributing editor at The American Prospect, where he was managing editor, he has written for The New York Times, the Financial Times, and Foreign Policy. He previously reported from Cairo and was a fellow at Harvard’s Radcliffe Institute. On X: @mideastXmidwest.

The post Claude and Grok Are at War in Iran, and Americans Have No Idea appeared first on The National Interest.