Ukraine is already fighting robot-against-robot battles: Australia admits it funds military AI faster than it governs it

By The Strategist (ASPI) · October 5, 2026.

On the Ukrainian front, "robot-versus-robot battles are already commonplace", writes analyst David Kirichenko in The Strategist, the commentary blog of the Australian institute ASPI. Ukraine is equipping its unmanned ground vehicles with AI-based target identification systems that detect enemy troops at greater distances and with more precision. Its medium-range drones —Hornet, Darts and RAM-2X— use autonomous terminal guidance, which according to the source makes them more resistant to electromagnetic warfare. Russia has responded with its Geran and Molniya drones, fitted with mesh networks and onboard AI for the same function, already deployed at scale according to the same analysis.

Ukrainian military planners, again according to Kirichenko, are already thinking about the next step: autonomous robots that clear enemy positions before the infantry advances, naval drones that wait at sea until activated remotely and then attack on their own, and AI that combines battlefield data to calculate optimal flight routes for long-range incursions inside Russia.

That pace is what worries Canberra. Analyst Malcolm Davis argues that the Australian Defence Force (ADF) is still betting on a small number of very expensive crewed platforms, a model he considers "out of step" with the speed of innovation in unmanned systems. His proposal is not full autonomy but semi-autonomous systems bought "cheaply, in volume and quickly" to add combat mass in a region of enormous distances. The underlying shift, in his view, is moving from keeping the human "in the loop" —piloting every decision— to keeping them "on the loop": supervising what the machine has already decided, instead of deciding for it.

That same logic is arriving underwater. Davis, together with Justin Bassi, recalls that Australia, the United Kingdom and the United States announced in May a project to equip autonomous submarines —such as the Ghost Shark, built by the company Anduril— with the sensors and weapons needed to navigate, detect and attack on their own. If the trio meets its 2027 delivery target —"a big if", the authors themselves qualify— the so-called Pillar Two of AUKUS will move from research to a real operational military capability. The underlying argument: Australia depends on 15 undersea cables for almost all of its internet traffic, and the most efficient way to protect them is with swarms of autonomous vehicles patrolling those routes, freeing crewed submarines to operate farther afield.

The problem, warns researcher Aina Turillazzi in another piece in the series, is that "AI is now funded in almost every line of the Australian defense budget" but "is not governed through them". Her diagnosis points to a risk already well known in the psychology of automation: "automation bias", whereby an operator under pressure stops evaluating the evidence and simply confirms what the machine has already decided. On top of that, according to Samuel White, comes a coalition problem: if Australia and its Indo-Pacific partners do not share the same criteria on when to require human verification before striking, they may share the same technology and still not be interoperable in what really matters. And analyst Ayeza Areej adds another layer: "continuous learning" systems —which keep updating with new data after deployment— are harder to predict and more vulnerable to adversarial deception. If one of them mistakes a civilian vehicle for a military target, responsibility is spread among the commander who authorized its use, the operator who activated it, the developer who designed it and the authority that certified it, with none of them bearing the full blame.

What this ASPI roundup portrays is not a distant military curiosity: it is the same pattern we have been observing in civilian AI, transplanted to a domain where mistakes cannot be undone. In our recent coverage of enterprise agent governance we noted that real authorization is built at the destination, not by whoever creates the system; here the equivalent is even more unsettling, because the authority to stop an autonomous weapon is diluted among whoever funds it, whoever operates it and whoever manufactures it, precisely when the pressure for speed is pushing the human out of the loop. That race —Ukraine and Russia perfecting terminal guidance, Australia buying capability before having clear rules— is, in the short term, a documented risk, not a science fiction scenario.

Over the longer term, however, the same technology that unsettles today is the one that could reduce human casualties in combat and protect critical infrastructure —like those 15 undersea cables— with cheap swarms instead of high-value assets exposed to a single strike. An abundance of low-cost autonomous systems, properly governed, could make defense less lethal for the people in uniform, not more lethal for everyone. But that outcome only arrives if governance stops trailing behind the budget, as Turillazzi warns. The lesson of this ASPI roundup is simple: the question is not whether military AI is advancing —it already is, daily, in the skies over Ukraine— but whether anyone is going to apply the brakes before it becomes necessary.