War at Machine Speed: How AI is Redefining Modern Warfare and Removing the Human Element

2026-10-07

Artificial Intelligence is fundamentally transforming warfare, shifting from human-controlled systems to autonomous operations at machine speed. This article explores how AI is dissolving traditional defenses, accelerating kill chains, multiplying mass effect, and automating targeting, raising critical questions about human control and future conflicts.

War at Machine Speed: When AI Inherits the Drone

The Last Meters and the Last Human

Somewhere over the contested ground near Sumy in the spring of 2025, a Russian strike drone lost its radio link and kept flying. There was nothing remarkable about the jamming; Ukrainian electronic warfare had been severing the umbilical cords of first-person-view drones for two years, and severing that link had become the defender's most reliable and cheapest form of protection. What was remarkable was that the drone did not fall, drift, or circle uselessly until its battery died. It had onboard a small computer-vision payload, recognized its target on its own, and finished the last stretch of the attack without a human hand on the controls and without any signal to interrupt (IEEE Spectrum, 2025). The brake that the defender had leaned on — cut the link, kill the strike — no longer worked, because in those final meters there was no link to cut.

That small event, multiplied across thousands of airframes and both sides of a war, is the hinge on which this article turns. The preceding transformations of contemporary warfare — the collapse of the joint kill chain into a single cheap platform, the arrival of what can be called precise mass in place of the few-and-exquisite munition, the hollowing of physical effect and its migration toward cognition, and the erosion of defense from a shield into a sponge — were all, at bottom, transformations that still ran through human beings. A soldier flew the drone. A cell of officers built the target list. A commander weighed a strike. Electronic warfare worked precisely because it attacked the one indispensable and irreducibly human element in the loop: the radio conversation between the operator and the machine. Artificial intelligence enters this system not as one more capability to be added to the ledger, but as the agent that removes the human from each of these places in turn — from the terminal meters of the strike, from the tempo of the kill chain, from the generation of mass effect, and from the act of defense itself.

The thesis of this article is therefore deliberately narrow and, I will argue, more unsettling for being narrow. Artificial intelligence does not add a new revolution on top of the revolutions already under way. It changes their nature by subtracting the human bottleneck — and, with it, the human brake. Where the cheap drone collapsed the kill chain into one platform, AI threatens to collapse the tempo of that chain to machine speed, faster than an adversary can perceive and respond. Where the drone made mass cheap, autonomy threatens to make mass self-directing, so that quantity no longer requires a proportionate quantity of operators. Where electronic warfare gave the defender a cheap last resort, machine-vision terminal guidance defeats it in precisely the meters where it used to be decisive. And where defense had already begun to fail as a coherent proposition, AI turns the encounter into a contest of two autonomous systems reacting to each other with no human in either loop — an AI-versus-AI duel fought below the threshold of human reaction time.

Yet the honest analyst must hold two ideas at once, and this article will insist on both throughout. The first is that the trajectory is real, fast, and already visible in the field, not merely in the laboratory or the vendor's brochure. The second is that it has not arrived. The most careful observers of the Ukrainian battlefield — the place where these technologies are being pressure-tested at industrial scale — report that genuine autonomy "is not yet present on the battlefield," that the "technology has not reached the required level," and that promises of an immediate revolution in autonomous drones are "premature" (Bondar, 2025; Kirichenko, 2025). First-deployment hit rates for autonomous seekers collapse to a third or half of the vendor's claim; troops under fire frequently prefer a wired fiber-optic drone that cannot be jammed to an autonomous one that cannot be trusted (Bondar, 2025). The story that follows is therefore not the story of a finished revolution but of a crossover point — the moment at which each of the older transformations changes its character because the last human is leaving the loop, told with careful attention to how far that departure has actually gone.

I proceed in nine movements. I begin with the hinge: the axis along which AI dissolves the defender's last cheap counter, electronic warfare. I then trace the removal of the human from the tempo of the kill chain (hyperwar), from the generation of mass (autonomous swarms), and from the industrial-scale act of targeting (AI targeting at Lavender and Maven scale). I turn next to the encounter that results when defense is itself automated — the autonomous duel — and then to the migration of all of this into the cognitive domain. I close with the honesty brake, an accounting of how much of this is arriving rather than arrived, and with the three questions on which the whole trajectory ultimately turns: the threshold of meaningful human control, the new asymmetry that autonomy may reintroduce, and the specter of a flash war fought faster than statecraft can intervene.

The Axis AI Dissolves: Electronic Warfare and the Terminal Meters

To understand why AI is a change in nature rather than degree, one must begin where the defender's economy of force had come to rest. Across two years of large-scale drone war, electronic warfare emerged as the single most cost-effective countermeasure available to a defending force. A drone that costs a few hundred dollars and carries a warhead capable of destroying an armored vehicle is an extraordinary bargain for the attacker only so long as it can be flown to its target. The radio link between operator and airframe is the vulnerable seam in that bargain. Jam the control frequency or the satellite-navigation signal, and the drone is reduced from a precision weapon to a falling object. Electronic warfare was, in the language of the offense-defense balance, the defender's last cheap brake — a countermeasure whose per-engagement cost was far below the cost of the thing it defeated, and which therefore promised, at least locally, to restore some of the advantage that cheap precision mass had stripped from the defense.

Machine-vision terminal guidance breaks that brake, and it does so in the specific stretch of the engagement where the brake mattered most. The principle is simple to state and consequential in effect: rather than depend on a continuous radio conversation all the way to impact, the drone carries a small onboard processor and a camera, and over the final few hundred meters it recognizes the target visually and steers itself, with no link to jam. The Ukrainian module known as TFL-1, produced by a company named for the science-fiction premise of a fourth law of robotics, is the emblematic case: a payload costing on the order of fifty dollars that assumes autonomous control over the last four to five hundred meters of flight and, in the developers' phrase, "ignores the radio horizon" — because in that final segment there is no radio dependence to exploit (The Defense Post, 2025; IEEE Spectrum, 2025). By 2025 such modules were reportedly shipping in the thousands (The Defense Post, 2025). The countermeasure that had defined the defender's economy is neutralized in the precise meters where it used to win.

The quantitative claims attached to this shift, if they hold, describe not an increment but a phase change in the value of the weapon. According to analysis of the Ukrainian field data, autonomous terminal navigation raises strike success rates from a range of roughly ten to twenty percent to something on the order of seventy to eighty percent; it cuts the effective cost of a successful strike by a factor of three to four, because far fewer drones are wasted per target destroyed; and it extends the range at which a target can be recognized and locked from around three hundred meters to roughly a kilometer (Bondar, 2025). Each of these numbers matters on its own, but their combination is what changes the strategic arithmetic. A weapon that lands four times as often, costs a quarter as much per kill, and can begin its terminal engagement three times farther out is not the same weapon with better tolerances; it is a different instrument occupying a different place in the offense-defense balance. And the industrial base to exploit it already exists: Ukraine is reported to have built on the order of two million drones in 2024 (Bondar, 2025). The scarce input is no longer airframes; it is the small measure of autonomy that lets each airframe finish its own attack.

Crucially, this is not a one-sided Ukrainian innovation that a symmetrical adversary lacks. The same logic has migrated to the Russian side of the line. Russian Shahed-type loitering munitions have been observed carrying Nvidia Jetson Orin modules and computer-vision payloads that allow autonomous navigation under GPS jamming, and the V2U system has been associated with AI-guided strikes in the Sumy area (IEEE Spectrum, 2025). This bilateral adoption is analytically important, because it forecloses the comforting interpretation that autonomy is merely a clever workaround one side found against another's electronic warfare. When both sides fit their strike drones with onboard vision, electronic warfare does not simply lose its edge against a particular opponent; it loses its status as a general-purpose cheap counter for anyone. The seam that the defender learned to attack — the human-machine radio link — is being sewn shut on every airframe at once.

This is the sense in which AI is a change in nature. The drone had already collapsed the joint kill chain into a single platform and made precise mass cheap; those were transformations that electronic warfare could still bite into, because they left a human in the loop and a link to the human that could be severed. By removing the human from the terminal meters, machine vision removes the seam. The defender's last cheap brake is not made weaker; it is made irrelevant in the segment where it worked. Everything that follows in this article is a variation on that single move — the human, who was simultaneously the system's bottleneck and its brake, is being taken out of one place in the loop after another.

Machine Speed: Hyperwar and the Compression of the OODA Loop

If the terminal meters are where AI dissolves the defender's cheap counter, the tempo of the kill chain is where it threatens to dissolve the defender's ability to react at all. The relevant concept is old enough to have a pedigree. The term "hyperwar" was coined by Allen and Husain in 2017 to describe a machine-speed, AI-fueled form of war in which the decision cycle is compressed beyond the pace of human cognition (Allen and Husain, 2017). The organizing frame beneath it is older still: John Boyd's observe-orient-decide-act cycle, the OODA loop, which holds that the side able to move through the cycle faster forces its opponent to react to situations that have already changed, until the slower side's decisions are perpetually addressed to a world that no longer exists.

The contemporary articulation of hyperwar makes the mechanism explicit. It is, in one recent formulation, "the combination of AI, autonomous systems, sensors, and software that compresses the OODA cycle to machine speed" — a system able to detect, decide, strike, assess, and disperse faster than an adversary can respond (Kasapoğlu, 2026). Each verb in that sequence corresponds to a human function that AI is now positioned to accelerate or assume: sensing fused and interpreted by machine rather than pored over by an analyst; deciding narrowed to a recommended course of action rather than deliberated from raw data; striking executed by a platform that no longer needs a continuous human hand; assessing performed automatically by the same sensors that cued the strike; and dispersing conducted before a counterstrike can be organized. When every step runs at the speed of software, the loop closes in seconds, and the human who used to sit at its center becomes the slowest and therefore the binding constraint.

The Ukrainian battle-management system Delta is the field instantiation of the front end of this loop. Delta fuses real-time intelligence, applies AI analysis, performs target identification, and coordinates strikes across a common operating picture, drawing the disparate feeds of a distributed war into a single machine-assisted decision surface (Kasapoğlu, 2026). It is precisely the kind of system in which the compression of the OODA loop stops being a theoretical claim and becomes an operational property: the interval between a target appearing to a sensor and a munition being cued against it shrinks toward the limit set by the machinery rather than by the meeting, the map, and the radio call.

Here, though, the honesty brake must be applied for the first of several times, and it is the same author who articulated the hyperwar thesis who supplies the caveat. AI is entering the kill chain today "mostly via limited functionality — assisted navigation, target-recognition, target-lock," with humans still supervising the consequential decisions (Kasapoğlu, 2026). That qualification is not a footnote; it is the whole difference between the trajectory and its destination. What exists now is a decision cycle in which AI has taken over the perceptual and preparatory functions — the observe and much of the orient — while the decide and, in most cases, the act remain nominally with a human supervisor. The compression is real and measurable, and it is precisely because the machine has already claimed the front of the loop that the human at the back of it has become the rate-limiting step. The strategic pressure this creates is obvious and dangerous: every second the human supervisor spends deliberating is a second of tempo surrendered to an adversary whose own supervisor may be deliberating less, or not at all. Hyperwar does not require anyone to decide to remove the human from the decision. It only requires that keeping the human in the decision become a competitive disadvantage. That is the mechanism by which a doctrine of human supervision erodes not by policy but by pace.

Autonomous Mass: Swarms and the Multiplication of Effect

The drone made mass cheap by making each unit cheap. Autonomy threatens to make mass self-directing by breaking the link between the number of drones in the air and the number of humans required to fly them. This is the third place from which AI removes the human — not from the terminal meters or the decision tempo, but from the span of control that has always limited how much mass a given force can actually wield at once.

The clearest current example is the Ukrainian company Swarmer, whose software has reportedly been used in more than seventy thousand operations and which reduces the number of operators required per mission from around nine to around three (Defense One, 2025). The reduction is the whole point. A force is not limited by how many drones it can buy but by how many it can meaningfully control; every operator is a human bottleneck, and every mission that needs nine of them can be run at only a third the scale of one that needs three. Swarmer's drones are described as able to exchange targets among themselves, choose payloads, and confirm damage — functions that, performed by the collective rather than by individual operators, begin to convert a group of separately piloted aircraft into something closer to a single distributed weapon (Defense One, 2025). The company's roughly fifteen-million-dollar Series A round in September 2025 is a modest signal that this reduction in the human-to-machine ratio is being treated as an investable proposition rather than a demonstration (Defense One, 2025).

The vision articulated at the frontier of this work makes the recursive logic vivid. Swarmer's founder describes a future of "mother drones" — "swarms of autonomous drones carrying other autonomous drones to protect them against autonomous drones trying to intercept them" (Azhnyuk, 2025). The sentence is worth dwelling on because it contains, in compressed form, the entire trajectory of this article: offense, defense, and counter-defense all autonomous, all reacting to one another, with the human present only as the distant author of the initial intent. If it were realized, mass would no longer be a quantity a commander marshals and directs; it would be a self-organizing swarm the commander releases.

But the honesty brake bites hard here, and it must. The gap between the vision and the reality is at its widest in exactly this domain. True cooperative behavior — many drones sensing one another, dividing tasks, and adapting jointly under fire — remains emergent rather than fielded. Most of what is marketed as "swarming" today is more accurately described as coordination software that reduces the operator count, not autonomous collective intelligence maneuvering as one under enemy action (Lieber Institute, 2024). The distinction matters both operationally and morally. Reducing nine operators to three is a genuine and significant multiplication of effect, but it is a quantitative easing of the human bottleneck, not its elimination; three humans still supervise, still choose, still bear responsibility. A genuinely autonomous swarm that selects and engages targets collectively raises a different and sharper problem, which the same source names precisely: the "accountability gap" (Lieber Institute, 2024). When a swarm acts as a collective and something goes wrong — the wrong target struck, the protected category killed — it becomes genuinely unclear who decided, and therefore who is answerable. The technology is arriving faster than the doctrine and the law that would locate responsibility within it. That gap is not a transitional inconvenience to be closed by better engineering; it is a structural consequence of removing the human from the span of control, and it will not resolve itself.

AI Targeting at Scale: From Lavender to Maven

The three preceding moves — dissolving electronic warfare, compressing the loop, multiplying mass — describe autonomy entering the tactical strike. The removal of the human from targeting at the operational and campaign scale is a different and in some ways more consequential development, because it concerns not the flight of the munition but the generation of the target list itself: the industrialization of the decision about what and whom to strike.

The Israeli system reported under the name Lavender is the most fully documented instance, and it should be read carefully rather than sensationally. According to the investigative account, Lavender marked on the order of thirty-seven thousand individuals as potential targets during the Gaza campaign, generating a target bank at a scale no human targeting cell could have produced by traditional means (Abraham, 2024). The reported human role in vetting those machine-generated designations was minimal to the point of vanishing: officers are described as spending roughly twenty seconds per target before authorizing it, with that human check often amounting to little more than confirming that the designated person was male (Abraham, 2024). The system carried a reported error rate on the order of ten percent, and strikes were reportedly authorized with pre-approved allowances of fifteen to twenty civilian deaths per junior target — parameters that convert the individual moral weight of each strike into a throughput setting (Abraham, 2024). A companion system reported as "The Gospel" performed the analogous function for structures rather than individuals, generating physical targets at machine scale (Abraham, 2024). One source's summary of the human contribution is the single most economical description of what it means to remove the human from targeting: "We only checked that the target was a man" (Abraham, 2024).

The American development points the same direction from a different starting position and carries the process one step further along the chain. Project Maven, which began as an effort to apply computer vision to drone footage, has matured into the Maven Smart System, developed with Palantir, which aggregates more than a hundred and seventy-nine intelligence, surveillance, and reconnaissance sources and applies computer vision together with large language models across the span from identification to strike (Allen and Mande, 2026). The scale claims attached to its recent use are striking: in the opening twenty-four hours of the June 2025 operation against Iran, the system reportedly supported strikes against more than a thousand targets — roughly a tenfold increase over prior tempo — with on the order of twenty personnel performing analytic work that had required more than two thousand during the Iraq campaign (Allen and Mande, 2026). The program was reported to be moving toward program-of-record status in March 2026 (Allen and Mande, 2026). The most revealing evidence, however, is not the numbers but the self-description offered from within the enterprise: "We've gone from identifying a target to now coming up with a course of action, to now actioning that target. All from one system" (Stanley, 2026).

That sentence deserves the same attention as the "mother drone" vision, because it marks the same threshold from the other side of the world. Identifying a target, devising a course of action against it, and executing that action were, in classical targeting doctrine, three distinct functions performed by distinct people at distinct moments, and the seams between them were where human judgment, legal review, and command responsibility entered. Collapsing all three into "one system" is the operational-scale analogue of the tactical collapse that machine-vision terminal guidance performed in the last four hundred meters. In both cases, the human seams — the places where deliberation, doubt, and accountability used to live — are engineered away in the name of speed and efficiency. The difference is one of altitude: the terminal seeker removes the human from a single strike; the targeting system removes the human from the manufacture of the strike list. When these two developments meet — machine-generated target banks feeding machine-guided munitions through a machine-compressed loop — the kill chain becomes an integrated machine from end to end, with the human present, if at all, as a twenty-second checkpoint whose only verified function was to confirm a category.

The Defensive Duel: When Both Sides Automate

There is a consoling intuition that the automation of the offense might be answered by the automation of the defense, and that the two might cancel to something like the old equilibrium. The evidence suggests the opposite: automating the defense does not restore the human's central place; it removes the human from the last function in which he remained — the act of defending — and produces an encounter in which two autonomous systems react to each other faster than any person in either chain can follow.

The defensive application of autonomy is already the most operationally mature part of this whole landscape, precisely because shooting down an incoming drone is a narrower and more permissible task than launching an offensive strike. Autonomous Shahed-hunters have moved from concept to attrition. Ukraine's MaXon and the US-linked Project Eagle, associated with a system reported as Merops, are credited with downing more than a thousand Shahed-type drones by November 2025 (IEEE Spectrum, 2025). At the AUSA exhibition in October 2025, Israel Aerospace Industries presented an autonomous command-and-control counter-unmanned-aircraft system that fuses electro-optical and radar sensing and coordinates multiple interceptors against each target at machine speed (IAI, 2025). The defensive kill chain, in other words, is being compressed to machine speed for the same reason as the offensive one: an incoming autonomous threat arriving faster than human reaction demands a defensive response faster than human reaction. The logic is symmetrical and self-reinforcing. Each side's automation of its own loop is the justification for the other's.

This is the AI-versus-AI duel, and its significance is that it removes the human from the last place he had been retained. In the offensive developments, one could still imagine — even where it was eroding — a human authorizing the strike. In an autonomous defense against autonomous attack, both the trigger and the counter-trigger belong to machines, because the engagement occurs below the threshold of human reaction time. To insist on a human in the loop of the defense is to accept being too slow to defend; to remove him is to accept an encounter no human is supervising in real time. The regulatory frameworks have quietly registered this reality. The updated United States policy on autonomy in weapon systems permits supervised autonomous systems to defend installations without co-located human operators, an explicit accommodation to the fact that point defense against machine-speed threats cannot wait for a human to close the loop (Scharre, 2023). Defense is the domain where the removal of the human is not a future risk but a present and formally sanctioned fact.

One further point sharpens the strategic stakes of the defensive duel, and it is a matter of uneven maturity rather than of principle. Western counter-unmanned-aircraft capability is assessed to lag that of Ukraine and Russia by something on the order of three to four years (Bondar, 2025). The states that have been fighting a large-scale drone war are, unsurprisingly, several development cycles ahead in the reflexive, autonomous defenses that such a war demands. This lag is a reminder that the autonomous duel is not a uniform global condition arriving everywhere at once but a capability distributed unevenly according to who has been forced to build it — a theme that will return, in a more consequential form, when we come to the question of the new asymmetry.

The Cognitive Front: Generative AI and the Migration of Effect

The transformations traced so far concern the physical strike — its terminal guidance, its tempo, its mass, its targeting, its defeat. But one strand of the broader story held that physical effect had begun to hollow out and migrate toward cognition: that against certain adversaries, blowing things up produces diminishing strategic returns, and that the contest increasingly runs through perception, belief, and will. It would be natural to expect generative AI, which produces persuasive text and imagery at near-zero marginal cost, to be the decisive accelerant of that migration — the instrument that finally makes influence at scale as cheap as precision mass made destruction at scale. The evidence to date offers a more instructive and more qualified picture, and it is precisely this qualification that makes the cognitive front the best-lit case study in the hype-versus-reality tension that runs through the whole subject.

The raw capability is real. In roughly ninety days in 2024, OpenAI disrupted five distinct state-linked influence operations that were using its models to generate mass multilingual content: the Russian operation known as Doppelganger, the Chinese operation Spamouflage, the Iranian operation associated with IUVM, and an Israeli operation identified as Stoic among them (OpenAI, 2024). Microsoft's Threat Analysis Center reported AI-augmented influence activity directed at the 2024 United States election (MTAC, 2024). Generative models plainly lower the cost and raise the fluency and volume of propaganda content across languages, which is exactly the input one would expect to see if the migration of effect toward cognition were about to be industrialized.

And yet the operations mostly failed at the thing that actually matters, and the reason they failed is the reason to be cautious about the entire hype cycle. As OpenAI's investigator observed, "these operations may be using new technology, but they're still struggling with the old problem of how to get people to fall for it" (Nimmo, 2024). Generation was never the binding constraint on influence; distribution and persuasion were. Producing a thousand fluent posts is trivial and always was relatively cheap; getting a real audience to see them, believe them, and act on them is the hard part, and generative AI does little to solve it. The bottleneck in cognitive warfare was never the writing. It was the credibility, the reach, and the human susceptibility on the other end — and those remain stubbornly resistant to automation.

The cognitive front thus functions as a controlled experiment in the central claim of this article, and its result cuts both ways with useful precision. Where AI removes a human bottleneck that was genuinely the binding constraint — the operator flying the drone through the last four hundred meters, the analyst poring over ISR feeds — the effect is transformative, because the thing that limited the system is exactly the thing removed. Where AI removes a bottleneck that was not the real constraint — the writer producing propaganda copy — the effect is marginal, because the actual limit lies elsewhere, in a human factor that no model touches. This is the discipline the subject demands. The removal of the human is decisive exactly and only where the human was the binding constraint, and the analyst's task is to distinguish the places where that is true from the places where it merely appears to be.