Military AI Challenges Human Accountability
Davit Khachatryan is an international lawyer and lecturer focusing on the intersection of armed conflict, emerging technologies, and international law.
Artificial intelligence is no longer confined to code-stained labs or military contractors’ slideshows: it has become a regular presence on modern battlefields. In 2024, as Israeli analysts relied on tools like Gospel and Lavender to generate targeting lists, the Pentagon set out to deploy swarms of autonomous drones through its Replicator Initiative. Targeting algorithms (AI systems analyzing data to identify and prioritize military targets) now compress the decision cycle from days to minutes, sometimes seconds, fundamentally challenging the way law, ethics, and accountability operate in armed conflict. It was in response to these realities that, on December 24, 2024, the UN General Assembly adopted Resolution 79/239, affirming that international humanitarian law (IHL) applies “throughout all stages of the life-cycle of artificial intelligence in the military domain” and calling for appropriate safeguards to keep human judgment and control at the heart of military decision-making.
But resolutions and declarations, while necessary, do not themselves restrain machines. The responsibility for lawful conduct must remain anchored in human actors: commanders, engineers, and political authorities. Algorithms, after all, have no legal personality; they cannot form intent, stand before a court, or bear the weight of tragedy or blame. This is why the real task for military commanders, policymakers, and legal advisers is about translating the timeless obligations of the laws of war into practices and workflows that keep the chain of accountability intact, even as machines accelerate the tempo of armed conflict beyond anything imagined by those who first wrote those rules.
The question, then, is whether states are willing and able to build safeguards so that, even as decisions speed up and control becomes diffuse, a human being remains at the end of every algorithmic chain of action.
What is a Military AI?
Ask three officers to describe what counts as AI in uniform, and you will likely hear three different answers. One will mention software that sorts satellite imagery, another will point to a drone that selects its flight path, and a third may describe a logistics program that determines which convoy moves first. All of them are correct because military AI is a broad spectrum of software-enabled capabilities that touch nearly every corner of modern operations.
At one end of this spectrum are decision-support algorithms. These tools sift through immense volumes of data and present patterns or anomalies for a human to review. They normally remain firmly subordinate to human choice; nothing happens until a commander or operator approves the recommendation. Further along are autonomous platforms that can steer themselves, prioritize targets, and in some cases, use weapons or make lethal decisions without direct oversight. Both adapt and learn from experience.
This capacity for continual learning is why military leaders are drawn to AI and also why lawyers are cautious. As these systems become more complex and more adaptive, it becomes significantly harder to demonstrate that every use of force still complies with IHL. The capacity to process information at machine speed promises new efficiencies and tactical advantages. It also threatens to outpace the ability of humans to scrutinize and override what the machine proposes. The more sophisticated the system, the greater the challenge of ensuring that its operation does not slip beyond the reach of law or human conscience.
Distinction, Proportionality, Precaution, and Indiscriminate Attack
At the heart of International Humanitarian Law are a handful of principles that every commander must observe in the conduct of hostilities, regardless of the technology at their disposal. The rule of distinction requires that attacks be directed at combatants and military objectives. This obligation reaches back to the design and validation of algorithms. If the data used to train a targeting model is biased, incomplete, or outdated, there is a real risk that the system will misclassify a school as an ammunition depot or mistake a civilian vehicle for a military convoy. Effective distinction, then, depends not only on rigorous data hygiene, continual red-team testing, and the capacity for the system to express its confidence in a way that human commanders can understand and question. When an algorithm cannot explain its reasoning, or when its output cannot be interrogated by a human, the legal requirement of distinction is at risk.
Proportionality is the next pillar. Even when a target is lawful, an attack is forbidden if the expected harm to civilians would be excessive compared to the anticipated military advantage. AI magnifies the challenge: it can process immense quantities of data and recommend actions at speeds that compress human decision-making to mere seconds. The speed can encourage a dangerous automation bias, where commanders are inclined to trust the machine’s judgment without fully weighing the consequences. To meet the proportionality requirement, there must be a clear, understandable record of what information the system used, what options were considered or rejected, and how the balance was struck between expected military advantage and potential civilian harm.
Precaution demands that every feasible step be taken to spare civilians and civilian objects before, during, and after an attack. In the context of AI, this means not only reviewing weapons before their use, but also conducting continual reassessment as software evolves or as new data and sensors are incorporated. This kind of legal review is often called an “Article 36 review,” referring to the provision of Additional Protocol I to the Geneva Conventions. That article requires states to determine whether any new weapon, means, or method of warfare they develop can be used consistently with international law. Even countries that are not parties to Additional Protocol I often conduct similar reviews. Effective precautions also require tamper-resistant records of every input and decision, so that when failures occur, they can be reconstructed and learned from. In uncertain conditions, systems must default to conservative modes, such as surveillance only, rather than risk unintended harm through automated action based on outdated or corrupted data.
Finally, IHL prohibits indiscriminate attacks. Any use of force that cannot be reliably confined to military objectives or that is likely to strike civilians and civilian objects without distinction is forbidden. AI promises new levels of precision, but it also presents new risks if boundaries are not clearly defined and tested. The only way to prevent this is through hard-wired limits on where and when systems may operate, relentless stress-testing under a wide variety of conditions, and the retention of a genuine human veto at every stage.
In sum, IHL principles remain in force, but their practical application now depends on embedding those rules into the very architecture and operation of AI.
Speed, Opacity, and Proliferation
AI exposes points of friction that IHL rules were never designed to anticipate. When AI is integrated into the targeting cycle or command and control networks, it can compress the decision-making process from hours to seconds. Early warning systems may communicate directly with automated defenses, and predictive algorithms can recommend preemptive action before any human has fully grasped the situation. The space for reflection, deliberation, and legal review narrows dramatically, and the traditional safeguards built around time for human intervention may vanish.
Opacity presents an equally serious challenge. Unlike conventional weapons or even most traditional software, AI often operates as a black box, producing outputs that even its creators cannot fully explain. When models are trained on synthetic or computer-generated data, or proprietary protections prevent independent scrutiny, the ability of lawyers and commanders to review, test, or question the system is severely limited. Under such conditions, the concept of a one-time weapons review loses much of its meaning, and the burden shifts to continuous, in-depth monitoring and oversight, tasks that are often difficult to sustain.
Proliferation is the third critical fault line. The trained neural weights, code, and the operating concepts of many military AI systems can be transmitted anywhere in the world in seconds. A commercial drone can be upgraded into a strike or reconnaissance platform with a software patch delivered by email, by repurposing its mission, or, if heavy enough, by using it as a primitive weapon, such as crashing it into a target. Thinking more creatively, these drones can be launched in coordinated swarms to overwhelm air defenses. As these technologies spread, and as militaries operate alongside coalition partners using different systems trained on different data and following different logic, the risk grows that the seams between systems will become legal blind spots.
Keeping Humans in Command
The purpose of IHL is to protect people and limit suffering during armed conflict. To achieve this, the law is written to make sure that responsibility for the use of force rests with human beings, not with machines. This cannot be maintained with the promise of meaningful human control as an abstract principle.
First, every new AI system deployed for military purposes must be subject to a recurring legal review. A record of these reviews, maintained transparently and with clear signatures from legal, technical, and operational authorities, would ensure that no system enters the field without documented human oversight and an unbroken chain of responsibility.
Second, the architecture of responsibility must run through the entire life cycle of each system. From the earliest stage of data collection and model training, through to deployment, field use, and after-action review, every layer should be linked to identifiable individuals or teams who are empowered to act and accountable for their choices. Embedding this kind of responsibility matrix within the metadata and operational logs of each system makes it possible to trace every decision and intervention back to its source, even years later, if a failure must be investigated.
Third, no AI tool should be fielded until it has been subjected to rigorous adversarial testing. Red-team exercises, designed to probe for bias, vulnerabilities, and failure modes, must become a routine part of military procurement and certification. Where deficiencies are found, the system should be withheld from deployment until those risks are resolved. This process must be a core responsibility of states and their armed forces.
Finally, safeguards must be built not only into the technology but into the very doctrine and organizational culture that guide its use. Preauthorized defensive systems can be kept within tightly defined geographic and temporal boundaries, while time-critical operations should still require streamlined but explicit human approval. Strategic or high-consequence strikes must always retain full deliberative review. Coalition operations and multinational partnerships need common standards and protocols so that interoperability does not become a backdoor for legal evasion. These measures are the living expression of the law’s demand that there is always a human face and a human name at the end of the chain of action.
Governing machines
AI is already changing the face of war. Its speed, scale, and adaptability have the potential to transform the conduct and the ethics of armed conflict, presenting risks as profound as its promised advantages. Yet, it is people, political leaders, military personnel, engineers, and lawyers who remain responsible for the choices that machines enable.
Resolution 79/239 is a clear assertion that IHL foundations must not be surrendered to the logic of the algorithm. The task ahead is not to demonize artificial intelligence, nor to place our hopes in technical fixes alone, but to ensure that the rules of war are translated into new domains and that the structures of oversight are robust enough to keep responsibility where it belongs.
If we succeed, AI may yet deliver on its promise of greater precision and restraint. If we fail, we risk allowing the tempo of technology to outrun the reach of law. In the end, the most important question is not what our machines can do, but whether we have the resolve and imagination to govern them, so that, even as the future unfolds at the speed of code, the chain of accountability remains unbroken.
