Introduction
On 6 July 2026, a Russian Molniya variant fixed-wing drone struck a petrol station in Zaporizhzhia city, tragically killing three civilians after apparently missing its intended target. The drone, operating as part of a swarm, had no radio antennas to communicate with a human operator and used the Nvidia Jetson Orin module – hardware that allowed the drone’s onboard vision model and flight controller to autonomously navigate, select its target, and carry out the strike.
The incident appears to be the first publicly documented case of civilian deaths attributed to a fully autonomous system that, once launched, could close its own kill chain. Reporting framed it as the arrival of a ‘dystopian’ future of war in which ‘killer robots roam the skies making life-or-death decisions.’ Col. Serhiy Minaiev, Zaporizhzhia’s air defence commander, similarly argued that soon, ‘we will be living in a Terminator movie… Machines are making decisions to strike’.
Terminator, however, is the wrong analogy as it suggests machines have assumed authority over warfare. In reality, while autonomous systems now exercise more operational control, humans still firmly retain command, setting the parameters within which these systems operate.
Are machines actually making decisions?
The references to Terminator and roaming killer robots are typical of the current debate on autonomous systems, which assumes that humans are outsourcing decision-making to machines and, as a result, losing control of warfare. This is misleading, however, and part of the problem is the word “decision.”
Certain systems can perform processes that could reasonably be described as decision-like. Arne Norlander, for example, writes of autonomous “cognitive systems”, which update internal models as environmental circumstances change and ‘decide which inputs are relevant to [their] continued operation.’
Nevertheless, the term “decide” warrants caution, as much of what current autonomous systems do is arguably closer to sophisticated stimulus-response than to human decision-making, which implies deliberation, contextualisation, and potentially even moral judgement.
For instance, AI-enabled uncrewed systems often use neural-network-based computer vision models for target detection and classification, including convolutional neural network (CNN)-based object detectors. A CNN classifier built by this author using the CIFAR-10 dataset illustrates the basic mechanism:
Author’s own CNN model using a simplified version of the CIFAR10 dataset.
This model did not deliberate. It was trained on thousands of labelled images, converting these into numerical representations of features such as pixel intensity and colour. These values were processed through successive network layers to output a probability score for how closely each image matched its trained categories.
Forensic evidence from the Zaporizhzhia strike suggests that propane tanks were the likely intended target category. If this is true, the system identified the target and acted without human intervention, likely taking a pre-defined course of action once its confidence score crossed a certain threshold.
Yet calling this a “decision to strike” implies a degree of independent agency. While the drone may have exercised control over its own movement and targeting, a human set the parameters within which it could act before launch.
Autonomy and the separation of Command from Control
The UK’s 2025 Strategic Defence Review noted that autonomy and AI are ‘changing the economics of Defence’, offering greater speed of decision, mass, and resilience to electronic warfare. These advantages, however, have implications for Command and Control.
Definitions of Command and Control are often ‘circular’ – NATO, for example, defines Command as the ‘authority … for the control of military forces.’ Even the acronym “C2” implies an innate connection between its parts. Yet, as autonomous systems proliferate, the concept of Command is increasingly distinct from Control.
Command is best understood as a human function rooted in ‘creativity and will’ and applied to the mission’s objective and operating parameters. Control, meanwhile, concerns the structures and processes through which the commander’s intent is translated into action, and risk is managed.
In a conventional remotely piloted aircraft, an operator retains control to alter course or select targets. In autonomous systems, that control passes increasingly to machines, while humans retain command over mission purpose and constraints.
This distinction was emphasised by Olli Pekka Suorsa, who argues that terms like “unmanned” or “remotely piloted” no longer capture how today’s drones function. He proposes the term ‘remotely commanded aircraft system’ to describe the shift from a human-in-the-loop who directly pilots the craft to a human-on-the-loop, who retains command but delegates control to a system for the management of its own flight, target acquisition, and strike.
The operational demands of modern warfare arguably necessitate this separation of Command from Control. Yet incidents like Zaporizhzhia have raised ethical concerns about deploying systems that are hard to control post-launch.
Command and Moral Control
Critics have compared autonomous drones to “fire-and-forget” weapons, where the ability to alter a weapon’s trajectory and point of impact is lost once fired. But control of an unguided munition is very different to the control an autonomous system can exercise over its own behaviour. A fire-and-forget munition cannot substantially reinterpret and respond to its environment, while an autonomous system can.
As David Omand argues, autonomous systems can be programmed to operate under an “adaptive moral control layer”, effectively formalising human moral authority in the system before deployment. Here, the human operator sets the Command parameters governing what the system is permitted to do – the characteristics of legitimate targets, acceptable collateral risk, and other constraints reflecting ethical judgements established before the mission.
This principle already exists in the weaponeering of precision-guided weapons – selecting a system and configuration to achieve an intended effect is itself a Command function, no different from selecting an autonomous system for a given mission. JDAM, for instance, allows different warhead configurations to be employed, including reduced-collateral-damage options such as the BLU-126/B.
This means that even with autonomous systems, humans still retain Command of these systems and, by hard-coding constraints into a system’s behaviour, can even extend that into a form of moral Control.
Conclusion
The Zaporizhzhia strike does not herald an era of “Terminator” warfare. More importantly, whether autonomous systems should be described as ‘deciding’ to strike obscures the more consequential question of how autonomy changes the relationship between Command and Control.
As autonomous systems proliferate across militaries worldwide, this separation of Command from Control will likely deepen, driven by the operational advantages of delegating Control to machines.
But this is not the same thing as humans losing control of warfare. Humans will be required to set the parameters governing these systems, potentially embedding a form of moral control into their behaviour.
And yet, in Zaporizhzhia, three civilians lost their lives in the strike. Open-source reporting attributes this to a navigation error, though whether that reflects a flaw in Command or simply the unpredictability of executing in a live environment is not publicly known.
This is a reminder that no matter how precisely the terms of engagement are set, war always retains the capacity for irreducible and tragic error.
