No Finger on the Trigger: AI and Warfare

Artificial intelligence has arrived on the battlefield faster than the laws designed to govern it and the gap between the two is widening every day. We are faced with a critical legal and ethical dilemma that does not seem to have a practical solution. To put it bluntly, when an algorithm selects a target, who carries the responsibility for the death that follows? Traditionally, the principles of international law, albeit hibernating in books and used selectively, would have been applicable in conflicts between states; however, the conundrum facing us is novel.

The laws of war were built around a deceptively simple assumption: that the person pulling the trigger could be held to account for it. The UN Charter governs when states may go to war; the Geneva Conventions govern how they must fight once they do. These two bodies of law operate independently. A war can be unlawful in its origins and still be bound by the rules of armed conflict once it begins. The symmetry is deliberate. Every belligerent, regardless of why they are fighting, is subject to the same prohibitions: no targeting civilians, no disproportionate force, no wilful destruction without military necessity. These rules were designed for human decision-makers such as soldiers who could receive orders, commanders who could be prosecuted and states that could be brought before tribunals.

However, international law was not designed to envision an answer to ‘the algorithm did it’. Accountability through the court structure was envisioned for a pre-algorithmic world. The current structure of accountability under international law is designed for human defendants and commanders, not for systems trained on data. An algorithm has no legal personality under the UN Charter making it impossible for a lawyer, commander or tribunal to interrogate and prosecute a system that made decisions in milliseconds. The regulation of how AI warfare is used cannot be bound by treaty obligations nor can it be prosecuted under the current legal order.

Currently no legal text or document covers substantive issues including the characterisation of autonomous weapons, preliminary legal considerations, proposed prohibitions and regulations and the application of international humanitarian law to such systems.

Autonomous Weapons

A Lethal Autonomous Weapons System, or LAWS, is essentially a weapon that can decide on its own whom to attack and when to fire, because no human needs to pull the trigger.

The word ‘autonomous’ covers a wide range of semi autonomous, supervised autonomous and fully autonomous weapon systems. At one extreme you have a drone that a human pilot controls remotely, approving every strike. At the other extreme you have a fully self-operating system that finds a target, identifies it and kills it entirely on its own, with no humans involved at any stage. Most of today's military AI sits somewhere in the middle of that range and that middle ground is precisely where governments and arms manufacturers are pushing hardest and fastest right now.

Algorithmic warfare does not factor in the risk of civilian casualties as it looks to enhance military efficiency and strategic advantage. The removal of human judgment and control increases the speed of violence and reduces humans to data points. A painfully real example of this can be found in the extensive use of AI enabled systems such as Lavender, Gospel and Where's Daddy used by the Israel Defence Forces. These AI-powered warfare tools are used for data collection, facial recognition, phone tracking, behavioural analytics, surveillance data, imagery, signals and infrastructure identification. Once Lavender, for instance, identifies a person deemed as hostile, it generates a kill list of suspected Hamas militants which is then treated as an order to hit without any human judgment, accuracy or check. More often than not, these suspected militants turn out to be women and children. Without a doubt, Gaza has become a testing ground for emerging disruptive warfare technology being used by the Israel Defence Forces leading to a demand in the global weapons market.

The proliferation of AI-enabled targeting capabilities has advanced considerably beyond the margins of a few technologically dominant states. No fewer than thirty nations are currently engaged in the development or active deployment of some form of algorithmically assisted weapons system. Among these, the United States, Russia, China, Israel, South Korea, the United Kingdom and Turkey have progressed furthest along the development curve. Yet despite this breadth of state participation in the militarisation of artificial intelligence, not one of these actors has accepted a legally binding commitment to prohibit fully autonomous lethal systems. The result is a structural asymmetry that sits at the heart of the current governance crisis: the technological trajectory is accelerating at a pace that existing diplomatic mechanisms — constrained as they are by consensus requirements, competing strategic interests and institutional inertia — have thus far proven wholly unable to match.

No Good Guys or Bad Guys

Anthropic, the tech giant responsible for Claude, was blacklisted by the Pentagon after it refused to grant unfettered access to its models for mass surveillance and fully autonomous weapons. They didn’t sit quietly; they took action by taking the Trump administration to court, challenging the blacklisting. Now on a bare perusal, I would cheer Anthropic for taking the moral high ground but there is more to the background story than we are exposed to. On 13th February this year, the Wall Street Journal reported that Claude was used during the January raid that captured Venezuelan leader Nicolás Maduro, deployed not directly, but through Anthropic's partnership with Palantir Technologies, whose platforms sit deep inside the United States Department of Defense. Reuters tried to independently verify the report and couldn’t.

Anthropic declined to comment on specific operations, the Pentagon declined to comment and Palantir said nothing. Now that silence tells you as much as any statement would have. The most telling detail, though, is this: sometime after the raid, an Anthropic employee reportedly asked someone at Palantir how Claude had actually been used in the operation. The company that built the model, and signed a contract reportedly worth $200 million, had to ask a third party what its own software did during a lethal strike on a foreign capital. That is not a company in control of its own ethical commitments, it is a company that sold access and lost visibility. There are no good guys here, and no bad guys either, just actors chasing the same advantage, each certain that stepping back means someone else steps forward.

Anthropic postures as the industry's conscience, while its models turn up in lethal operations it cannot fully account for. The Pentagon claims to sever ties while still leaning on the tools it claims to have dropped. The accountability gap this piece keeps circling back to isn't a gap between bad actors and good law. It's a gap that every actor gets to hide inside, good or bad, because nobody involved has to answer for what the machine actually did. There is a rush and surge to acquire the most lethal AI warfare, and in this pursuit companies are blurring the lines. Even the appearance of ethical restraint deserves scrutiny.

Tribunal for AI Warfare Crimes

History has taught us one lesson: international legal bodies capable of adjudicating mass atrocities have invariably emerged as reactive rather than preventive. The International Criminal Court, conceived as a permanent remedy to the inadequacies of ad hoc justice, was nonetheless the product of decades of failure rather than foresight. The Nuremberg and Tokyo trials, the International Criminal Tribunal for the former Yugoslavia (ICTY) and the International Criminal Tribunal for Rwanda (ICTR) are all by-products of where the law arrived late, shaped by the contours of the devastation it sought to address rather than designed to forestall it.

The challenge confronting us now is whether we disrupt this pattern or continue with our head down until we are standing knee deep in mutual destruction. The rapid militarisation of artificial intelligence presents an opportunity, perhaps the last available to construct an adequate legal and institutional framework before, rather than after. Whether states possess the political will to act prospectively, rather than waiting for the familiar logic of genocide and consequence to reassert itself, remains the defining question of this moment in the development of international humanitarian law.

What's Next?

In the absence of a binding international legal instrument governing the use of artificial intelligence in armed conflict, domestic regulatory frameworks have emerged as an imperfect and inherently limited substitute. A number of states have enacted, or are in the process of developing, national legislative frameworks addressing the deployment of AI systems in governmental decision-making contexts. The European Union's Artificial Intelligence Act, which entered into force in 2024, represents the most comprehensive of these efforts to date, building a tiered system that grades AI applications by risk and demands greater scrutiny the higher the stakes: transparency requirements, human oversight and proof of accuracy before deployment.

However, the Act's reach is fundamentally undermined by an explicit carve-out for AI systems deployed in the service of national security. This exemption is not merely a technical limitation, it is a fundamental contradiction at the heart of the EU's regulatory ambition. It means that the precise domain in which algorithmic decision-making carries the gravest consequences for human life and international stability, military and intelligence applications, is the domain that the framework deliberately and expressly leaves ungoverned. The practical effect is to replicate at the domestic level the same governance vacuum that persists at the international level: the more consequential the AI application, the less likely it is to be subject to meaningful legal constraint.

In December 2024, the United Nations General Assembly adopted a resolution on lethal autonomous weapons system with a support of 166 votes in favour, 3 opposing and 15 abstentions. Similarly in 2024, the Group of Governmental Experts on lethal autonomous weapons systems, chaired by Ambassador Robert in den Bosch of the Netherlands, introduced a significant procedural development in the form of a rolling text, a collaborative working document through which participating states attempt to build consensus around language that might form the basis of a future regulatory instrument on autonomous weapons systems.

However, it would be prudent to not overstate the strength of the resolution and the working document. They represent, at best, a framework for negotiation rather than a commitment to one. This perhaps serves as a reflection of where diplomatic consensus currently stands, useful as a record of shared concerns and emerging areas of agreement, but far short of the binding instrument the situation demands. The votes are there, the language is inching forward. What is missing, as ever, is the will to act before catastrophe makes the choice for us.