On February 28 hundreds of missiles, guided bombs, and autonomous drones streaked toward targets across Iran in a joint American-Israeli offensive code-named Operation Epic Fury and Operation Roaring Lion. Within the first twelve hours, allied forces had executed nearly 900 strikes — and artificial intelligence was embedded in nearly every layer of that effort.
It was a moment that humanity is unprepared for which crept up on us without proper debate: a major war in which AI-assisted targeting, autonomous drone swarms, and machine-speed decision cycles were not experimental features but operational realities.
Sources familiar with the U.S. military’s use of artificial intelligence tell CBS News that AI programs — including one created by Anthropic, which the Trump administration has deemed a supply chain risk — are likely being deployed as part of the U.S. operation against Iran. On March 20, the Pentagon announced further use of Palantir’s AI tool, Maven, even though it incorporates Anthropic’s software, Claude.
While the Pentagon has not said exactly how AI tools are being deployed, CBS News spoke with several experts with knowledge of military operations who described the likely scenarios.
“The military is now processing roughly a thousand potential targets a day and striking the majority of them, with turnaround time for the next strike potentially under four hours,” retired Navy Admiral Mark Montgomery, senior director of the Foundation for Defense of Democracies’ Center on Cyber and Technology Innovation, is quoted as saying. “A human is still in the loop, but AI is doing the work that used to take days of analysis — and doing it at a scale no previous campaign has matched.” This is because AI can ‘read/interpret’ everything all at the same time in a way which would take hundreds or thousands of humans to match.
That scale is precisely what troubles experts. Craig Jones, senior lecturer in political geography at Newcastle University and author of The War Lawyers, describes what is happening as a compression of the “kill chain” — the sequence of identifying, approving, and striking targets. “You’re reducing a massive human workload of tens of thousands of hours into seconds and minutes,” Jones explained. “You’re reducing workflows, and you’re automating human-made targeting decisions in ways which open up all kinds of problematic legal, ethical and political questions.”
Machines That Kill
Lethal Autonomous Weapons (LAWs) are military systems that identify, select, and engage targets without human intervention and can be aggressive or defensive. For years there have been discussions of a ban on such weapons at the United Nations but nothing has come of those talks. U.S. military policy does not ban the use of LAWs but DODD 3000.09 requires that all systems, including LAWs, be designed to “allow commanders and operators to exercise appropriate levels of human judgment over the use of force.”
With good reason. A recent article in The New Scientist detailed how researchers at King’s College put frontier models into 21 simulated nuclear crises—and in 95% of them, the models treated nuclear escalation as a strategic move rather than a last resort.
This is because an AI model learns the structure of a strategic argument without the deeper human and societal context in which that argument was meant to operate, LinkedIn co-founder and venture capitalist Reid Hoffman says in a March 11 post.
He points out that reasoning is not the same as judgment. “A model can generate a compelling explanation for escalation, and still be missing the most important variables,” argues Hoffman. “Mercy. Context. Ambiguity. The recognition that de-escalation is the true form of strength. That the smartest move is not always the most aggressive one. That the actual objective is not dominance — it’s survival.”
The public debate about AI and defense has been too shallow, he says in the post, which was published amid an escalating conflict between Anthropic, the AI company behind the chatbot Claude, and the U.S. military. The conflict centers on Anthropic’s push for guardrails that would explicitly prevent the military from using Claude to power fully autonomous weapons or conduct mass surveillance on Americans. The Pentagon, for its part, demanded the ability to use Claude for “all lawful purposes,” so that the U.S. can keep pace with other countries, like China, that are using AI.
“If an AI company says its models are not ready for autonomous lethal decision-making, that should not be waved away as squeamishness or lack of patriotism,” says Hoffman. “In tech, the bias is usually toward claiming the product can do more. So, when the people closest to the models say ‘No, this system is not ready for that level of authority’ we should hear the warning.”
The Illusion of Human Control
Hoffman makes the argument that putting humans in the loop pays off. He points to Stanislav Petrov, a Russian lieutenant colonel of the Soviet Air Defense Forces who played a key role in the 1983 Soviet nuclear false alarm incident: in 1983, when an early-warning system indicated incoming U.S. missiles, notes Hoffman, Petrov made a human judgment — that it made no sense for the U.S. to be attacking — and decided not to fire his missiles back, even when all the machine indicators said otherwise.
It makes sense that the riskier the setting in which powerful AI systems are deployed, the more we turn to the idea that humans should always be the ones to make the final decisions, says Financial Times columnist Sarah O’Connor in a March 17 article in The Financial Times. In the context of war, the public and regulatory debate (and one source of the recent row between Anthropic and the U.S. government) has focused on the seemingly binary distinction between fully autonomous weapons and those that are subject to “human control.” In the corporate world, too, the deployment of semi-autonomous agents has led companies to turn to experienced humans as the ultimate decision-makers.
But, she asks, does this solution necessarily lead to the best of both worlds, in which machines boost speed, accuracy and productivity, while humans supply expertise, context, judgment and accountability?
The answer, in my opinion, is no.
As O’Connor points out, AI operates at superhuman speed. On the battlefield, for example, even systems that leave final decisions to humans can churn through mountains of data and vastly increase the number of potential targets to hit. But when so-called “kill chains” are compressed from hours to minutes or even seconds, it calls into question how much real-time control humans can realistically provide.
There is a human toll when AI agents vastly speed up the pace and volume of work that still needs to be directed and reviewed by humans. UC Berkeley professors Aruna Ranganathan and Xingqi Maggie Ye of the Haas School of Business ran an eight-month study (April–December 2025) observing around 200 employees at an undisclosed U.S. tech company that provided free access to enterprise generative AI tools. The study, which was published in The Harvard Business Review in February, found that workload creep led to cognitive fatigue, burnout, and weakened decision-making, and that the initial productivity surge gave way to lower-quality work, turnover, and other problems.
Another issue is that many humans are inclined to trust machines even when they are warned not to. The phenomenon of “automation bias” – the propensity for humans to favor suggestions from automated decision-making systems and to ignore contradictory information made without automation, even if it is correct – is well documented.
“There is instinctive appeal to the idea that humans must have the final say over these powerful new technologies, but the history of human-machine interaction tells us that is not as easy to achieve as it sounds,” writes O’Connor. “What’s more, the illusion of human control can be more dangerous than its clear absence.”
Blurring Lines
At the core of the debate is a question that no legal framework yet adequately answers: if an autonomous system makes a mistake and kills civilians, who is responsible? The pilot who launched the drone? The software engineer who wrote the targeting algorithm? The general who approved its deployment? Or no one?
There has been much speculation about how AI is being used to help identify targets. The February 28 strike that hit a girls’ school in Iran, killing 168 people, many of them children, is a case in point. The U.S. is believed to bear responsibility. If the data the AI was using was out of date, then the killing of schoolchildren could be attributed to AI because it identified the target as within a military base. The same error could have been made by humans, and shifting blame to an AI system could be useful to avoid taking responsibility. There is also a third possibility: a human could wrongly be obliged to take responsibility for something they had no control over. Academic Madeleine Clare Elish introduced the concept of a moral crumple zone to describe how responsibility for an action may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system. Just as the crumple zone in a car is designed to absorb the force of impact in a crash, the human in a highly complex and automated system may become simply a component—accidentally or intentionally—that bears the brunt of the moral and legal responsibilities when the overall system malfunctions. While the crumple zone in a car is meant to protect the human driver, the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator.
A Turning Point With No Map
Against that backdrop, U.S. Secretary of Defense Pete Hegseth said that he wanted everyone in the military to use AI, explaining that “at the click of a button, AI models on GenAI can be used to conduct deep research, format documents and even analyze video or imagery at unprecedented speeds.” It is unclear how much AI training every member of the military has had. And, as businesses are beginning to discover, it is one thing to train people how to use AI systems technically and another to train them how to use those systems wisely.
Important decisions need to be made:
- How often and when should a human review an AI decision, and what data can the machine provide for the human to make an evaluation?
- How do we ensure that humans have a clear understanding that they are interacting with a machine? An estimated 20% of American men report that they have an AI ‘intimate partner’ and 72% of teens consider their AI chatbot to be a friend. It is easy to imagine that these people would be willing to trust what the AI tells them without question.
- What should AI training look like? The military requires pilots to put in many hours of training in their aircraft and simulators. Should all users of AI in the military be given extensive training to ensure that the humans in the loop can perform their role?
- How often should humans rotate out of the job or shift? Checking whether an AI is doing something right is very boring work, especially if it usually gets it right. Ironically, creating systems that hallucinate less is making it more difficult to use humans in the loop effectively.
- Should the military be required to inform elected officials when and how AI systems have been used?
There is an opportunity for the military to develop sound AI governance principles, testing procedures, proper training, and recommendations on how to regulate AI in warfare. But there is no time to waste. Until the right legal, political and ethical frameworks are put in place, the concept of humans in the loop is an illusion which allows AI that is not fit for purpose to be used in the most dangerous circumstances.
The technology has moved faster than the law, faster than ethics, and faster than training courses and democratic oversight. There are no international treaties governing autonomous weapons. There is no agreed definition of what it means for a human to remain “in the loop.” There is no liability framework for an algorithm that kills the wrong people. The Iran conflict has made it undeniably clear that AI-powered war is no longer a future scenario – it is the present reality. Unless we act quickly, the AI-powered killing machines we built to win wars will prove to be a turning point with no map.
Kay Firth-Butterfield, one of the world’s foremost experts on AI governance, is the founder and CEO of Good Tech Advisory and the author of Co-existing with AI: Work, Love and Play in a Changing World (Wiley). Until recently she was Head of Artificial Intelligence and a member of the Executive Committee at the World Economic Forum. Last February she won The Time100 Impact Award for her work on responsible AI governance. Firth-Butterfield is a barrister, former judge and professor, technologist and entrepreneur, and vice-Chair of The IEEE Global Initiative for Ethical Considerations in Artificial Intelligence and Autonomous Systems. She was part of the group which met at Asilomar to create the Asilomar AI Ethical Principles, is a member of the Polaris Council for the Government Accountability Office (USA), the Advisory Board for UNESCO International Research Centre on AI, ADI and AI4All. She sits on the board of EarthSpecies and regularly speaks to international audiences addressing many aspects of the beneficial and challenging technical, economic, and social changes arising from the use of AI. This article is part of a series of exclusive columns that she is writing for The Innovator.
This article is content that would normally only be available to subscribers. Become a subscriber to see what you have been missing
