The U.S.-Israel war against Iran entered its third week amid political confusion and mixed messaging. The Trump regime initially told the press the attacks would last around four weeks. On March 9, Trump told CBS News the war was “very complete,” contradicting his own Department of Defense which almost simultaneously said, “This is just the beginning.”
In the days since, the DOD messaging has been the prevailing one, with bombings continuing to escalate despite Trump claiming March 13 the war will end “any time I want it to” and “when I feel it in my bones.”
Artificial intelligence has played a key role in the Trump regime’s propaganda campaign, both to engage with the populace and bend the perception of reality to be more favorable. But underlying that relatively simple use of AI as a content producer is something more insidious.
The same AI responsible for what the DHS Assistant Secretary for Public Affairs Tricia McLaughlin called “banger memes” also is partially responsible for dropping bombs on Iran.
AI goes to war
The U.S. military’s Maven Smart System is part of an initiative launched in mid-2017 under the auspices of Deputy Secretary of Defense Robert O. Work. Then known as the Algorithmic Warfare Cross-Functional Team — or AWCFT — the objective of the project was to analyze the volume of data available to the Department of Defense at a scale and speed impossible for humans. In particular, the first task of AWCFT was to develop and deploy machine-learning tools that could scan drone footage, detect and classify objects, and alert human analysts to important events.
In early December 2017, a mere six months from its inception, the project’s first algorithms were deployed to support drone missions against ISIS.
The Department of Defense quickly realized it could not develop and deploy AI systems on its own and turned to Silicon Valley for support. Soon, Amazon and the CIA joined forces with chipmaker Nvidia and satellite operator DigitalGlobe to begin building AI-powered tracking technology using satellite imagery that would eventually be rolled into Project Maven. Google partnered with the Pentagon to allow its AI systems to assist in video analysis. Other companies followed: Amazon AWS, IBM, Microsoft, Sierra Nevada Corporation, CrowdAI, Palantir.
Under the auspices of Palantir, Project Maven evolved from applying computer-vision algorithms on drone and satellite imagery into a broader decision-support platform that allowed AI systems to be part of the decision-making process in assigning weapons and executing strikes.
In 2020, the XVIII Airborne Corps led Project Maven through its first live-fire exercise. In a controlled experiment at Fort Liberty, an AI system used satellite imagery to locate and identify a tank as a hostile threat. It relayed this information to a human analyst who approved the strike. Having received approval, the AI then signaled a rocket launcher to strike the tank based on its satellite coordinates. It was the first time American soldiers struck a target located and identified by an AI system.
An AG3 Labs quadrocopter drone flies armed with a pink dummy bomb after it launched from an unmanned ground vehicle (UGV) during a demonstration of capabilities at the xTech Edge Strike: Ground robotics technology evaluation of the U.S. Army at the Grafenwoehr training grounds on March 13 near Grafenwoehr, Germany. The U.S. Army 2nd Cavalry Regiment has been hosting the competition, in which soldiers have been testing UGVs from a variety of manufacturers in realistic scenarios since March 3 to assess future procurement options of ground-based drones for the U.S. Army. (Photo by Sean Gallup/Getty Images)
War at the speed of AI
Traditionally, the process of determining a strike was methodical and labor intensive, involving a number of humans in collaboration with one another utilizing a variety of resources. Known as a kill chain, this process moves step by step to find, fix, track, target, engage and assess targets. Twenty years ago, during Operation Iraqi Freedom, it took a team of 2,000 staffers 45 minutes to work through the process. That was heralded as the most efficient in U.S. military history.
Project Maven’s first test took more than 12 hours. But four years and 10 iterations later, Maven had managed to match the efficiency and workload of OIF’s 2,000 staffers with a crew of only 20. That speed was not enough. With those successes, military leaders hoped to be able to make one thousand high-quality battlefield decisions in just a one-hour period. What used to take days or hours would now — through AI-assistance — take only seconds.
In the first 12 hours of its attack on Iran, the U.S. military launched strikes on almost 900 targets. They were able to do this because, Maven — running on Amazon Web Services and using a version of Anthropic’s Claude — was able to suggest targets and issue precise coordinates. Of the six steps in the kill chain, Maven is involved in every one except for engagement. It is finding targets, assessing them and making recommendations for which weapons to use in a strike.
“Of the six steps in the kill chain, Maven is involved in every one except for engagement.”
Humans still push the button. But the information guiding that decision is gathered, filtered and explained by AI systems. The U.S. military has become, in the words of Secretary of Defense Pete Hegseth, “an AI-first warfighting force.”
Craig Jones, a senior lecturer in political geography and expert on kill chains, told The Guardian, “The AI machine is making recommendations for what to target, which is actually much quicker in some ways than the speed of thought. So you’ve got scale and you’ve got speed … That might have taken days or weeks in historic wars. (Now) you’re doing everything at once.”
This overwhelming show of force — in which the U.S. spent more than $11 billion in six days — is precisely what the Trump regime views as strength.
The problem with AI judgment
The problem with AI is that it sometimes gets things wrong. We all know this.
Anyone who ever has used an AI chatbot knows it can be a people-pleasing yes-man that hallucinates research and gets basic facts incorrect. The same is true with military-grade AI.
Eric Schmidt, former chairman of Google and the Defense Innovation Board, has said the military needed to get used to “trying things, failing quick, all the standard Silicon Valley kind of things.”
Jane Pinelis, who oversaw testing of Maven in its early days, said the U.S. military needed to increase its risk tolerance.
Katrina Manson, author of the forthcoming book Project Maven, writes, “Perfection simply wasn’t possible: There could be errors resulting from AI hallucinations, faulty data and the way algorithms tend to lose accuracy over time.”
The human element — the person responsible for actually triggering the strike — is supposed to be the mitigating factor. It’s up to them to evaluate what the AI is saying. But large-language models have a way of being persuasive. Because humans have not done the work of sifting through the information, they often lack the critical context or personal investment needed to properly vet that information.
And when those decisions must be made at an increasing rate of speed, it becomes easy to simply rubber stamp the AI recommendations.
Elke Schwartz, professor of political theory at Queen Mary University in London, suggests the sheer quantity of targets leads to automation and action bias — the rapid machine decisions force human decisions to become automatic and mechanical.
Further, the more humans rely on artificial intelligence, the less intelligent they become. A recent study from MIT used EEGs to record brain activity in three groups of essay writers. The first group were allowed to use AI, the second allowed a regular internet search engine and the third allowed only to use their brains. The study found AI users consistently underperformed at neural, linguistic and behavioral levels. They had less ownership and understanding of the material they generated.
“When individuals fail to critically engage with a subject, their writing might become biased and superficial,” the authors said. “This pattern reflects the accumulation of cognitive debt, a condition in which repeated reliance on external systems like LLMs replaces the effortful cognitive processes required for independent thinking.”
“AI will not only make mistakes, it will cause human decision-makers to become more mistake-prone as well.”
Continued use of AI decision-making in military usage could very well lead to the same. The human decision-makers begin to think less critically, feel more detached from their work and begin to make biased, superficial and flawed results.
AI will not only make mistakes, it will cause human decision-makers to become more mistake-prone as well.
When the kill chain fails
One of those 900 initial strikes on Iran targeted a set of buildings that had once been part of an Islamic Revolutionary Guard Corps complex. Ten years ago, those buildings were a naval barracks. Today, they comprised the Shajareh Tayyebeh primary school. Using outdated information, the United States military sent a Tomahawk missile into an Iranian school, killing 175 people, most of them children.
It is unclear if Maven played any direct role in selecting the school as a target. It appears possible Maven listed the school as a target based on historical data, given that the U.S. has developed plans and targets for potential war with Iran for decades. Given the speed of the attack, human vetting may have missed that the compound was no longer part of the military complex.
“Anybody who thinks AI is going to magically solve the fog and friction of war is lying to you.”
Speaking to the Washington Post, now-retired Air Force Lt. Gen. Jack Shanahan said: “I’ve been pretty vocal since early Maven days that feeding current and accurate data into your model is your biggest challenge. As the tempo of the war increases, and the pressure to find more targets increases, there have to be checks and balances in place to ensure the targets being nominated for strike are legitimate targets, and you can catch any mistakes that might lead to civilian casualties and collateral damage. … It’s tragic this happened. And it shouldn’t happen again. And anybody who thinks AI is going to magically solve the fog and friction of war is lying to you.”
When Google’s connections to Project Maven came to light in 2018, it led to massive pushback within the company and Google choosing not to renew its contract in 2019. Former Google engineer Laura Nolan warned then, “If we are not careful, one or more of these weapons, these killer robots, could accidentally start a flash war, destroy a nuclear power station and cause mass atrocities.”
Fully autonomous weaponry is a clear and present threat. But just as threatening is a reliance on artificial intelligence that removes the resources of actual intelligence. Human oversight should have seen that Shajareh Tayyebeh was a school. Publicly available satellite imagery made it obvious.
Artificial intelligence promises to remove human limitations from warfare. But those limitations — our very humanity — always has been one of the few things slowing down the violence. The more we compress the kill chain, the less room remains for reflection, doubt and restraint. War may always be inhuman, but handing it over to machines risks making it even less human than before.
Amid the inhumanity of war, do we really need to lose even more of our humanity?
Josh Olds
Josh Olds is a public theologian and pastor for those disillusioned with institutional church. He is the creator of the small-group video series “Year on the Mountaintop” and a featured contributor to Fostering Hope: A Prayerbook for Fostering and Adoptive Parents. Follow his work on Facebook or at JoshOlds.com.
Related articles:
Springtime for war crimes | Opinion by Ginny Brown Daniel
Donald Trump has been wrong about most everything | Analysis by Rodney Kennedy
White House wins gold medal in internet fakes | Analysis by Josh Olds