Symposium on Cyber, International Law and Artificial Intelligence in Armed Conflict: Military AI and Common Article 1 – From Accountability to Governance

[Marco Roscini is Professor of International Law at the University of Westminster and a Senior Fellow at the NATO Cooperative Cyber Defence Centre of Excellence]

Most discussions of military AI and international humanitarian law (IHL) focus on autonomous weapons, targeting decisions and accountability. AI does raise difficult questions concerning compliance with the rules on the conduct of hostilities and the allocation of responsibility when AI-enabled systems contribute to violations of international law. AI, however, also triggers broader questions that arise well before an attack takes place and continue after it has ended. Decisions concerning the design, development, testing, procurement and deployment of AI systems may ultimately determine whether those systems can be employed consistently with IHL.

Common Article 1 of the 1949 Geneva Conventions offers a useful yet underexplored framework for examining these issues. Under this provision, States undertake “to respect and to ensure respect” for the Geneva Conventions “in all circumstances”. The obligation has received considerable attention in relation to arms transfers, third-State obligations and duty to exercise influence over coalition partners. Its implications for military AI, however, have received little consideration. This is surprising because, as already said, many of the choices that affect compliance with IHL are made long before AI systems are employed during military operations.

AI is relevant to Common Article 1 in two distinct ways. First, it may assist States in respecting and ensuring respect for IHL. More specifically, by processing information on a scale and at a speed unattainable for human operators, AI can support target identification, collateral damage estimation, military training and legal advice. Secondly, AI is itself subject to the regulatory scope of Common Article 1. States employing AI-enabled systems remain under an obligation not only to respect IHL, but also to ensure respect whenever those systems are designed, developed, tested, acquired, deployed or used by others in connection with an armed conflict.

This second dimension is the focus of this post. Properly interpreted, the obligation to ensure respect is not confined to reacting once violations have been committed. It also requires States to take all reasonable measures to prevent foreseeable violations and minimise the risks associated with the use of AI in the military domain. Compliance with IHL, therefore, depends not only on decisions taken during military operations, but also on choices made throughout the entire lifecycle of AI systems. Common Article 1 offers a framework for examining the governance of military AI from its design and development to its operational use.

Due Diligence Across the AI Lifecycle

Understanding the obligation to ensure respect as a due diligence obligation has important consequences for the military uses of AI. States are not required to guarantee that violations of IHL will never occur. Rather, as already said, they are expected to take all reasonable measures to prevent foreseeable violations and minimise the risks associated with the military uses of AI. The question, therefore, is not only whether AI complies with IHL once it is employed on the battlefield, but also what measures States are expected to adopt throughout the lifecycle of AI systems to reduce the likelihood of violations.

The obligation to ensure respect first applies to a State’s own organs and to those individuals whose conduct is otherwise attributable to it. AI may assist States in discharging this obligation in several ways, for instance by improving situational awareness, facilitating target identification, supporting collateral damage estimation, enhancing military training and assisting military legal advisers. At the same time, if States choose to rely on AI systems in the conduct of military operations by their organs or agents, they remain under an obligation to ensure that those systems can be employed consistently with IHL. This has several important consequences.

First, military personnel cannot reasonably be expected to employ AI consistently with IHL unless they understand both its capabilities and its limitations. Education and training, therefore, assume particular significance in this context. Training should be specific to the military functions for which AI is employed, while military manuals and rules of engagement should provide guidance on how to conduct AI-enabled operations consistently with IHL.

Secondly, AI systems need to be properly tested and submitted to legal review. Article 36 of Additional Protocol I already requires States Parties to review new weapons, means and methods of warfare. Common Article 1 reinforces this obligation. In the case of AI, in particular, legal review cannot be regarded as a one-off exercise. Unlike many conventional weapons, AI systems may evolve through software updates, retraining or changes in the operational environment. Its performance also depends on the quality of the data on which it relies. Testing and review, therefore, must necessarily be a continuing exercise throughout the operational life of the system. The same applies to the mitigation of algorithmic bias and protection against cyber vulnerabilities capable of affecting compliance with IHL.

Finally, ensuring respect for IHL by State organs and agents also requires an appropriate institutional framework. States need to adopt legislation addressing AI-enabled violations of IHL, clarify the applicable modes of liability and establish effective procedures for investigating incidents involving AI systems. Accountability remains important, but its function is not only to allocate responsibility after violations have occurred. It also contributes to preventing and deterring violations by strengthening compliance throughout the entire lifecycle of military AI.

Technology Companies and the Governance of Military AI

Common Article 1 also requires States to ensure respect for IHL by private individuals and legal persons on their territory or under their jurisdiction. This aspect is particularly significant in the context of military AI because of the significant role played by technology companies in designing, developing and maintaining AI systems. Many of the decisions that determine whether AI can ultimately be employed consistently with IHL, therefore, are taken well before an armed conflict begins and by actors that are not themselves parties to it.

States are thus expected to adopt reasonable measures to ensure that tech companies do not contribute to violations of IHL. The obligation is, again, one of due diligence. It does not require States to guarantee that private actors will always comply with IHL, but it does require them to regulate the design and development of military AI in a manner that reduces foreseeable risks. This obligation can be fulfilled through a range of different measures.

First, compliance with IHL should be taken into account from the earliest stages of the AI lifecycle. States should adopt legislation requiring designers and developers of military AI to conduct IHL assessments and to incorporate compliance-by-design requirements into the process. This necessarily requires close cooperation between legal advisers, engineers and software developers so that the principles of distinction, proportionality and precautions are considered before AI systems are deployed. Developers should also be required to identify and mitigate foreseeable risks arising from inaccurate training data, algorithmic bias and cyber vulnerabilities.

Secondly, States should make greater use of public procurement to ensure compliance with IHL. Procurement contracts should require suppliers to demonstrate that AI systems have undergone appropriate testing and legal review, that foreseeable risks have been identified and mitigated, and that systems continue to be monitored after deployment. Seen from this perspective, procurement may become an important indirect means through which States discharge their obligation to ensure respect for IHL by anyone under their jurisdiction.

Thirdly, military users need sufficient information about the capabilities and limitations of AI systems. Human oversight cannot be meaningful if operators are unable to understand how AI has produced a certain output or the circumstances in which it may be unreliable. This is particularly relevant because AI-enabled decision-support systems may trigger automation bias, leading operators to place undue reliance on AI recommendations instead of exercising autonomous judgment. States, therefore, should require developers to provide adequate information concerning system capabilities, limitations and foreseeable risks.

Finally, the obligation to ensure respect also extends to the transfer of military AI technologies. Before authorising the export of AI systems or related technologies by tech companies, States must assess whether there is a foreseeable risk that they will be used to commit violations of IHL. The obligation requires States to undertake an appropriate assessment before allowing the transfer of technologies capable of increasing humanitarian risks.

The External Dimension of Common Article 1

Whether Common Article 1 requires States to ensure respect for IHL only by persons on their territory or under their jurisdiction, or also by parties to an armed conflict beyond their jurisdiction, remains controversial. A growing body of scholarship (see for instance here and here), the International Committee of the Red Cross (ICRC), several States, the UN Security Council and the International Court of Justice have interpreted the obligation as extending beyond a State’s own territory and jurisdiction, requiring States to exercise whatever influence they possess to prevent or bring to an end violations of IHL. Others, however, reject this interpretation. It is not my purpose to revisit that debate here. Assuming that Common Article 1 also has an external dimension, this dimension can play an important role in military AI.

First, AI may assist States in ensuring respect for IHL by the parties to an armed conflict. Indeed, AI can strengthen the monitoring of compliance with IHL by processing satellite imagery, open-source intelligence and other digital information capable of identifying patterns of violations. This may enable States to identify situations in which diplomatic or other lawful measures may be appropriate to prevent further violations or encourage compliance with IHL by the parties to an armed conflict.

At the same time, States must refrain from encouraging violations of the Geneva Conventions and must not contribute to such violations by providing support where there is a foreseeable risk that it will be used inconsistently with IHL. Applied to military AI, this means that States must not facilitate AI-enabled violations by transferring AI systems, technical expertise, operational data or intelligence where there is a foreseeable risk that they will be used in breach of IHL. Before authorising the transfer of military AI technologies to a party to an armed conflict, therefore, States must assess that risk. As elsewhere under Common Article 1, the obligation is one of due diligence: it requires States to take all reasonable measures available in the circumstances rather than to guarantee that violations will not occur.

The external dimension of Common Article 1 also has another aspect: States are expected to exercise whatever influence they have, through lawful means, to prevent or bring to an end AI-enabled violations of IHL committed by the parties to an armed conflict. Depending on the circumstances, this may include diplomatic pressure, action through international organisations or other lawful measures capable of promoting compliance with IHL. What is required is not the achievement of a particular result, but the exercise of the degree of influence that may reasonably be expected in the circumstances of each case.

Military AI strengthens the practical significance of this aspect of Common Article 1. AI systems are increasingly transferred between States and integrated into military cooperation. Decisions concerning technological assistance, intelligence sharing and defence partnerships, therefore, may have direct consequences for compliance with IHL. If Common Article 1 is interpreted as also having an external dimension, therefore, States must consider not only how they employ AI themselves, but also whether their support for, or cooperation with, other States may contribute to AI-enabled violations of IHL.

This post forms part of the Opinio Juris symposium on International Law and Artificial Intelligence in Armed Conflict (introduced here) and draws on the author’s chapter in the forthcoming OUP volume of the same title