The Latest Chapter In American Military Muddling

John Arquilla is a long-time defense analyst and author of “The Troubled American Way of War” (Polity, 2026). On day one of America’s supposedly “most precise” aerial bombing campaign in history, 150 people, including 120 children, at Iran’s Shajareh Tayyebeh Elementary School, were killed. Over the following weeks, nearly two dozen more schools along with 10 hospitals were also hit by air strikes. And a strike on a petroleum facility near Tehran caused a persistent “oily rain” to fall on heavily populated areas, leading to a range of immediate and still-emerging health problems. These incidents and others like them — bombs falling on homes and public spaces — undermine U.S. Secretary of Defense Pete Hegseth’s claims of precision. But whatever the level of accuracy, or lack thereof, of the joint U.S.-Israel military actions, it is clear that the mistakes made during “Epic Fury” are being laid at the door of artificial intelligence; in particular, the Maven Smart System, an AI platform that is supposed to integrate nine major sensory systems — including video, signals intelligence and radar — and draws upon historical databases as well as the “real-time” information of drones, satellites and other technologies. The tragedy at Shajareh Tayyebeh was blamed on AI’s use of old data from before the school opened a decade ago; the site sits next to an Iranian Revolutionary Guard Corps facility. Human beings remained “in the loop” regarding the decision to strike, but no one took the time to check the latest satellite feeds to reverify the old data. The result? A terrible man-machine failure. Why did this incident and so many others like it occur? The simplest, most accurate answer is the U.S. Air Force’s obsession with high operating tempo. AI-enabled targeting is coveted because a bot can wade through massive tidal flows of information, integrating and assessing every bit of data at a pace far beyond human capacity. The appeal is a faster turnaround from detection to weapons release, and more accurate destruction done via this “kill chain.” One of the U.S. Air Force’s more extravagant acronyms to describe this tenet is F2T2EA — find, fix, track, target, engage and assess. The end goal is for AI to enable the U.S. to hit 1,000 targets per hour, according to the Air Force’s “Dragon” initiative. The Army has its own concept for using AI to speed up its missile, drone and artillery kill-chain rate to such a speed: Scarlet Dragon. If ever achieved by either service, this would mean the average time spent on any strike decision would be under four seconds. In the intense, sustained bombing of the opening weeks of the air war on Iran that began Feb. 28, the “Dragon” goal remained a distant dream, as the average number of targets hit each day hovered around 500. In theory, this equates to each target being contemplated and analyzed for just under three minutes. A long time for AI, but still a terrible rush for the “human in the loop,” who is supposed to be giving final approval for any given strike. It should come as no surprise, then, that more schools and so many hospitals were hit. Nor is it surprising that the Israeli bombing of Gaza, guided to a significant degree by AI (for example, by the recommendations of its Habsora, “the Gospel” bot), led to the killing of tens of thousands of innocents, a significant proportion of them children. Clearly, the propensity of humans “in the loop” with final decision-making authority is to defer to the algorithm-based judgment of the bot charged with integrating all sensor data to inform and guide air strikes against often fleeting targets. And what is true of the Israel Defense Forces holds as well for its close American ally. Indeed, a kind of shared strategic culture has been developing in recent decades, with ideas particularly about the impact of AI on military affairs being imported and exported between the two. The seminal U.S. Air Force doctrinal document on AI, “Note 25-1, Artificial Intelligence,”was released just last year. It contains the prediction that “AI will supercharge Intelligence, Surveillance, and Reconnaissance (ISR) by providing networked sensors capable of identifying hidden ‘needles in a haystack.’” But in addition to this expression of unbridled enthusiasm, there is also some awareness of the fresh risks that AI brings: “Inaccurate data sets can ‘poison’ AI systems with bad data, leading to erroneous decisions.” This is a prescient point, given what unfolded in Iran. Yet, for all its thoughtfulness, this official U.S. Air Force doctrinal document has very little to say about actual battle doctrine, offering instead merely rhetorical phrases like “global precision attack” and “rapid global mobility.” What this document is missing is a discussion of the issue that has bedeviled the Air Force since its inception as an independent armed service nearly 80 years ago: the emphasis on using air power against enemy homelands, despite the fact that few strategic bombing campaigns have achieved victory on their own. The first post-World War II effort to achieve what University of Chicago scholar Robert Pape has labeled the goal of “bombing to win” arose during the Korean War, when Pyongyang and other cities in North Korea were flattened and many hundreds of thousands of civilians — some estimates range up to nearly a million — were killed. But that conflict ended in a bloody draw. In the next war, the Air Force dropped more bomb tonnage on North Vietnam than fell on Nazi Germany. Yet the Vietnam War ended in ignominious defeat for the United States. The list goes on. For example, the shock-and-awe bombings of Iraqi cities and infrastructure, first in 1991 then again in 2003, did little to aid the advance of the invading ground forces, which were maneuvering quite well largely on their own. And in that second war with Iraq, strategic air power utterly failed to prevent the rise of nettlesome terrorist and insurgent campaigns that ran on for years. Even the seeming successes of American strategic air power are not absolute. For example, the 78-day bombing campaign against Serbia in 1999 did contribute to the willingness of Slobodan Milosevic to withdraw his troops from Kosovo. But the actual damage done to concealed Serbian ground forces was negligible. And in the autumn of 2001 the Taliban held out stubbornly during several weeks of U.S. strategic bombing. It was only when air power was at last repurposed from targeting Afghanistan’s sparse infrastructure to providing close support to the Green Beret-led teams of friendly local fighters that the Taliban were toppled — at least for that moment. Over the next 20 years, however, Taliban fighters figured out how to survive strategic air attacks on their havens in the tribal areas of Pakistan and throughout the areas of Afghanistan where they were reclaiming control. Their swift overthrow of the American-supported regime in Kabul during the summer of 2021 was little slowed by the significant air power assets the U.S. military had left the Afghans. Then-Chairman of the Joint Chiefs of Staff, Gen. Mark Milley, testifying to Congress a month after the debacle in Afghanistan, could only admit that for all its military prowess, the United States had suffered a stinging “strategic defeat.” AI True-Believers: From Washington To Kyiv The hope of Air Force generals and senior policymakers today is that AI will overcome this historic inability of air power alone to win wars by improving our strategic bombing. The vision of all-conquering American air power was first conjured by Gen. Billy Mitchell, who, in 1925, argued: “Aircraft possess the most powerful weapons ever devised.” He continued: “An air force using explosive bombs and gas [!] may cause the complete evacuation and cessation of industry.” The air war on Iran has done nothing to bring Mitchell’s dream closer to being realized. It is ironic that Mitchell was court-martialed for his unflagging advocacy of air power — the actual charge against him was “insubordination” — as he was soon elevated to sainthood, if not outright martyrdom, by a long procession of generals starting during World War II and continuing into the present. A pity that the trial judges, the youngest of whom was Mitchell’s friend Douglas MacArthur, didn’t call for any testimony to rebut the defendant’s ideas about the potential of strategic bombing, for this was much needed. Instead, the judges’ panel focused solely on his most heated, over-the-top rhetoric, in which he charged both the War and Navy Departments with an “almost treasonable administration” of American military and security affairs, earning him the insubordination charge. A similar charge could be levied against senior military leadership today, though for reasons exactly opposite those that animated Mitchell in his crusade for strategic air power. Now, the indictment should be for these leaders’ mindless devotion to Mitchell’s belief in strategic bombing to win. Secretary Hegseth’s boasting about the efficacy of the air attacks on Iran — which clearly failed to cow Tehran into submission — should be Exhibit A in this new indictment. With Pentagon head Hegseth now playing the role Mitchell pioneered a century ago. Air Force generals and others in the Pentagon — in and out of uniform — have told me they hope AI will provide a magic solution to the problem of strategic bombing’s inefficacy. But it won’t, as the evidence from air attacks on Iran, Gaza and Lebanon proves. Yet AI can do several other things for air power. The most important is in the use of autonomous drones for close support in tactical land and naval operations. This has been demonstrated in the Ukraine war, where drones now dominate the battlespaces on the ground and in the Black Sea. And since the Ukrainians make and deploy thousands more drones than they have trained drone operators, they are now often using a human operator to oversee the guidance of several drones into the general battle area and identify targets, before having AI take over completely for the final “strike phase.” AI is also used to counter Russian jamming, with the bot taking over the attack when the radio link to the drone operator is blocked. Such operations show we are getting close to a watershed moment that points to a future of war waged by intelligent machines. What the Ukrainians are pioneering is truly a quantum leap in tactical operations; these are their first steps in fighting with weapons meant to operate without the human in the loop. The success of these early Ukrainian uses of situational autonomy for drones in combat heralds a transformation in military affairs. It limns a future in which AI-piloted drones won’t need a radio link to a controller; thus, they will be much less vulnerable to enemy electronic jamming. They will be able to operate in coordinated “swarms” that mount simultaneous, omnidirectional attacks — a relatively straightforward programming matter — with a speed and accuracy far greater than any group of human drone operators could ever hope to achieve. This future can already be glimpsed as, over the past year, the Ukrainian military has scaled up its AI drone capacities, helping stymie the far more numerous attacking Russian troops, who have gained ground at a lesser daily rate than even the British soldiers did in the disastrous Somme offensive a century ago during World War I. Broader Military Applications For AI The range of AI applications in air power doesn’t stop at small drones, for it is already possible to replace human pilots in fighter plane cockpits. Indeed, as early as August 2020, an AI named “Falco” defeated an elite pilot who graduated from the Air Force Weapons School’s F-16 Weapons Instructor Course in five dogfights waged on F-16 Raptor simulators. Falco was far faster and more accurate in every round, emerging unscathed. The AI’s advantage grew from its ability to learn from practice, the process of so-called “machine learning,” which has been dramatically demonstrated by AIs defeating human world champions at the games of chess and Go. In the case of dogfighting, Falco prepared for the match over a short period by studying the equivalent of roughly 12 years’ worth of training flights and practicing one-on-one air duels against itself and other programs. An AI does not become bored or sloppy with the endless repetition of such training. It just gets better. And better. In addition to mano-a-mano dogfights, perhaps the most important organizational concept is integrating fully autonomous fighter planes into air squadrons alongside human pilots. The current vision is to introduce groups of robotic “collaborative combat aircraft” (CCA) into fighter and bomber formations. In the swarming aerial battles that many believe will characterize future air wars, a technologically advanced enemy may also employ AI-piloted craft; the CCA can engage opposing bots more effectively than humans. Indeed, a robot wingman may become the most sought-after commodity in future air warfare. Beyond individual dogfighting and larger-scale variations of aerial combat, AI already demonstrates superb skill across a range of activities central to the broader “air enterprise.” But AI has especially amazing potential to better manage the vital intelligence, surveillance and reconnaissance (ISR) that air platforms have long provided to terrestrial forces at the tactical level. An example that resonated widely across the military occurred during a Fort Benning wargame in 2019, when a small defense force faced a far larger attacking formation. The small force had ISR managed entirely by AI, while the big attacking force’s tactical use of ISR was controlled by human decision-makers. The small force soundly defeated the larger attackers. Research and experimentation continued along those lines and led to this year’s launch of the Robotic Autonomous Systems Leader Tactics (RASLT) course at Fort Benning. It may seem counterintuitive that the Army’s on-the-ground tactical forces would be making the best use of AI, but the Air Force’s focus on AI for strategic bombing is less impactful than simply having AI help humans fight better. The U.S. Navy is just as aggressive as the Army in its pursuit of AI-based enhancement of tactical capabilities, with this year’s fleet exercises off Key West using AI to run both ISR and fully autonomous aerial drone swarms. AI In Supporting Roles For all its hard-wired fixation on how AI can effectively wrestle with the chimera of strategic bombing, the Air Force has pursued an interesting range of other uses for AI. For example, it turns out that AI does spectacularly well at figuring out when equipment is most likely to need repairs, retrofitting or other forms of care. This “predictive maintenance” function not only extends the life of aging planes and other tech, saving money and potentially lives, but it also makes the Air Force more battle-ready. The Air Force has also made good use of AI for its logistics. Since air campaigns may be waged at great distances from home bases, like in the Iran war, logistical support is crucial to mission success. This includes functions like aerial refueling and consistently stocking forward operating sites so that their weapons inventories can meet the demands of the campaign. It turns out that AI can handle this sort of tasking extremely well. Then there is the whole matter of training human service members. Here, too, much Air Force training is being handled now with meticulous care by AI, with human performance carefully tracked and graded. To varying degrees, the Army, Navy and Marines are also using AI in these supportive ways. Finally, all the services have realized that hackers use AI to penetrate critical information systems, and so a broad consensus has emerged to grant AI the freedom to mount cyber defenses. Automation & Nuclear Strategy Senior U.S. officials and military leaders have been using AI to tackle high-level national strategy and crisis decision-making for years, with basic programs in use since 2017, but the process has accelerated rapidly this year. Because of their roles in any crisis involving possible nuclear escalation, the Air Force and Space Force have played major roles in American efforts to integrate AI into their risk analyses. Here, unfortunately, the news is not good. There is little training data for AIs to study, as only two atomic bombs have ever been detonated in wartime, and both by the United States. Threats of their use have generally been vague and ill-conceived, from President Dwight Eisenhower’s nuclear saber-rattling against China in the 1950s to Russian President Vladimir Putin’s not-so-veiled threats to use “small” nuclear weapons in the Ukraine war. The alternative to relying on the historical record is for AI to proceed with “machine learning” by playing countless wargames against itself and other AI programs. Much of this has been going on, and the results are startling and deeply troubling. A seminal study conducted by Kenneth Payne, a strategic studies professor at King’s College London, employed Claude, Gemini and GPT 5.2 in 21 crisis simulations and found that, as Payne put it, “Nuclear escalation was near-universal; 95% of games saw tactical nuclear use, and 76% reached strategic nuclear threats.” The problem: AI has no sense of moral abhorrence about waging nuclear war. Given that Russia still has an automated “dead hand” (aka “Perimeter”) nuclear command and control system, set to operate if Putin and his minions have all been killed, the findings from the King’s College study, as well as from other, less public, experiments, which also found that AIs were quick to push the nuclear button, are cause for great concern. This is especially the case given hypersonic and increasingly AI-smart missiles that can outfox any defenses, even President Donald Trump’s proposed Golden Dome. Preventing AI-automated escalation to nuclear warfare must be a goal of any future nuclear arms control agreements. The simplest solution would be for all nuclear powers to agree to a policy of “no first use” (NFU) of nuclear weapons, and to a covenant that NFU would be hard-wired into all AI connected to any systems that could authorize delivery of nuclear warheads. Verification would be a complex process, but no more so than the process that applies to another type of behavior-based agreement, the Chemical Weapons Convention. As to other military uses of AI, particularly in air power, there are ways to reduce the risk of error, malfeasance and ineffective and morally reprehensible results. The most important solution is to realize that not all automation needs to be at “artificial general intelligence” levels, that is, a level of computational power that exceeds human calculational abilities. As far back as the late 1980s, the philosopher of science Roger Penrose argued that AI could become quite clever at chess and other calculation-driven activities and would be highly useful in limited ways. But he believed that AI would never achieve the insight, moral judgment and sensitivity to context and innate knowledge that comes from human consciousness. The problems Claude, Gemini and GPT 5.2 have had with nuclear scenarios attest to Penrose’s prescience. Thus, the most important lesson for the future of AI, in and beyond military affairs, may be to optimize the use of its limited incarnations and to impose limits on efforts to optimize the development of the most advanced forms of artificial intelligence. The Pope Offers Hope Any effort to slow the pace of growth in cutting-edge artificial general intelligence must be undertaken under the auspices of international agreement. Otherwise, any nation’s unilateral effort to restrain AI development will undoubtedly be exploited by one or more rivals who see the chance to leap ahead in what is clearly becoming an AI arms race. The only prudent way ahead, then, is to apply the arms control concept to AI — selectively drawing upon how such limitations have been applied to nuclear, biological and chemical weaponry in the past. A good starting point for the discourse can be found in Pope Leo XIV’s encyclical, “Magnifica Humanitas,” released in May. The bottom line for the pope is to prevent the rise of lethal weapons that take human beings out of the loop. Certainly this point can be agreed upon by all when it comes to weapons of mass destruction. And surely also in cases of strategic bombing of civilian infrastructures. Perhaps even in the realm of anti-satellite weapons. And when it comes to areas where AI can function nonlethally in defensive ways, like protecting cyberspace from mass disruptive hack attacks, naval fleets from armed drone swarms and civilian infrastructure from being crippled by advanced computer viruses, it seems that even the pope’s view of limited use of AI in military settings still allows room for its development and automated use to protect life and the systems that sustain and enrich it. It is now our collective duty to find a way to continue the discourse the pope has prompted in a productive manner that speaks to the prospect of a future less replete with peril and more likely to foster peace and broad prosperity.