How Should Military AI Be Evaluated Under IHL?
Artificial intelligence (AI) is becoming an integral part of military operations. AI-enabled decision-support systems (AI-DSS) support intelligence analysis, target verification, operational planning, and the execution of military operations by collecting, integrating, and analyzing vast amounts of information beyond unaided human capacity. As these systems become more deeply integrated into military decision-making, they raise an important legal question: how should military AI be evaluated under international humanitarian law (IHL)?
The legal and ethical debate has largely answered this question by focusing on the risks AI may pose to compliance with IHL. Significant attention has been devoted to accountability, meaningful human control, automation bias, opacity, distinction, and proportionality. These concerns are well founded and remain indispensable. Yet evaluating military AI primarily through the possibility that it may undermine compliance with IHL captures only part of the law’s regulatory framework. IHL not only prohibits or restricts forms of conduct but also requires parties to an armed conflict to take affirmative measures to protect civilians. Central to this protective framework is Article 57 of Additional Protocol I, which requires those who plan or decide upon an attack to take all feasible precautions. Although recent scholars have begun to examine the relationship between AI and these obligations, the potential contribution of AI-DSS to their implementation has received less systematic attention than the risks associated with their use. This imbalance is particularly significant because AI-DSS are designed to improve the collection, integration, and analysis of data on which precautionary decisions depend.
This post argues that military AI should be evaluated not only by the risks it poses but also by its capacity to assist commanders in implementing IHL’s existing precautionary obligations. Reframing the debate in this way does not diminish the importance of the humanitarian risks associated with military AI. Rather, it places those risks within a broader legal framework that assesses not only how AI may undermine civilian protection but also how, under appropriate conditions, it may strengthen it.
The discussion develops in three stages. It first explains why the principle of precaution provides the appropriate legal framework for evaluating military AI. It then argues that the appropriate standard of comparison is not AI against an idealized human decision-maker, but AI-assisted and unaided human decision-making under comparable operational conditions. Finally, it argues that a complete humanitarian assessment of military AI must extend beyond individual attacks to consider its broader operational implications.
Viewing AI Through Precaution
The principle of precaution provides the appropriate legal framework for evaluating military AI, directing attention not only to whether an attack is lawful but also to the process by which that determination is reached and implemented. Article 57 of Additional Protocol I requires constant care to spare civilians and civilian objects throughout military operations. It also requires those who plan or decide upon an attack to verify that the intended target is a lawful military objective and to take all feasible precautions in the choice of means and methods of attack. It further requires that an attack be canceled or suspended if it becomes apparent that the target is not a military objective or that the attack would otherwise be unlawful. Compliance therefore depends not only on the final decision itself, but also on the information reasonably available to commanders and the procedures through which that information is collected, assessed, and acted upon.
Renato Wolf has begun to elaborate this connection between the principle of precaution and AI-DSS in the context of target verification. He divides the verification process into three stages: the legal qualification of potential targets, their classification, and their subsequent identification and location. He argues that precautions to verify must be applied throughout these stages and that measures specific to AI-DSS, including precautions outside the immediate control of the human operator, may sometimes be necessary to reduce false positives and ensure compliance with IHL. His analysis demonstrates that applying IHL to AI-DSS cannot be confined to reviewing an algorithmic recommendation at the final point of human judgment. The system’s operation throughout the target-verification process may itself affect whether the required precautions have been taken. Recent commentary by Jessica Dorsey and Luke Moffett similarly emphasizes that AI is becoming embedded throughout the targeting process rather than functioning merely at its final stage. Although their concern is that this development may increase humanitarian risks, their analysis likewise illustrates that the legal assessment of AI must extend beyond the final attack decision to the processes through which targets are identified, verified, and evaluated.
Wolf’s analysis also has a broader implication. Although the obligations set out in Article 57 remain unchanged, what constitutes a feasible precaution depends on the operational and technological circumstances in which an attack is planned and conducted. An AI capability that materially improves target verification or reduces false positives may therefore alter what precautions are considered feasible. This does not mean that every available AI capability automatically becomes required. However, where AI materially reduces the risk of erroneous targeting without imposing excessive military costs, its use may become necessary to satisfy the obligation to take all feasible precautions.
Jessica Dorsey approaches precautions from a different angle, arguing that the duty of constant care and the requirement to take feasible precautions should be understood as prior protective obligations. On this view, proportionality operates as a final safeguard, applied only after all feasible steps to avoid or minimize civilian harm have been taken, and some incidental harm remains unavoidable. This sequencing does not make targeting a rigid or linear process, as precautions must still be continuously reassessed in light of changing circumstances, including during the proportionality analysis itself. Her argument nonetheless highlights a key limitation of relying on proportionality alone in assessing AI-DSS: before weighing incidental harm against anticipated military advantage, commanders must first exhaust all feasible measures to prevent or reduce that harm.
Taken together, these analyses demonstrate why the principle of precaution provides the appropriate legal framework for evaluating military AI. Wolf demonstrates that AI-DSS may become relevant to the implementation of precautionary obligations throughout target verification. Dorsey explains why those obligations cannot be displaced by an exclusive focus on proportionality. Yet, evaluating military AI solely in terms of the risks it poses to compliance with IHL provides an incomplete basis for legal assessment. A complete assessment must also consider whether AI enables commanders to implement precautionary obligations more effectively. That assessment, however, requires determining the appropriate basis for comparing AI-assisted and unaided human decision-making.
The Appropriate Standard of Comparison
Current debates on military AI frequently assess these systems by comparing their performance with that of human decision-makers. Much of this discussion implicitly measures AI against an idealized standard of human performance. Although this comparison is understandable, it is not the standard of comparison required by IHL.
The principle of precaution does not regulate military operations under ideal conditions. Article 57 of Additional Protocol I requires those who plan or decide upon an attack to take all feasible precautions in the circumstances ruling at the time. Military decisions are therefore understood to be made under conditions of uncertainty, incomplete information, time pressure, stress, and rapidly changing operational circumstances. The legal obligation is not to achieve perfect outcomes, but to take all feasible measures to reduce the risk of civilian harm under those circumstances.
This understanding has important implications for evaluating AI-DSS. If precaution is inherently contextual and feasibility-based, the relevant comparison is not between AI and an idealized human decision-maker, but between AI-assisted and unaided human decision-making under the same operational conditions. The relevant legal question is not whether AI achieves perfect accuracy, but whether it enables commanders to implement precautionary obligations more effectively than would otherwise be possible.
This comparison also reflects the purpose for which AI-DSS are designed. Unlike autonomous weapon systems, AI-DSS are intended to support rather than replace human judgment. By integrating intelligence from multiple sources, analyzing large volumes of information, identifying relevant patterns, and continuously updating operational assessments, these systems may improve target verification. They may also facilitate the identification of civilian presence, support the selection of means and methods of attack that reduce incidental civilian harm, and enable commanders to respond more rapidly when changing operational circumstances require an attack to be modified or canceled. As experts observe, “Data analysts are no longer capable of manually scrutinizing all the raw data generated from reports, sensors, videos, and other data feeds.”
Importantly, the legal obligation to take all feasible precautions remains with the commander, who retains responsibility for the final judgment as to whether an attack may proceed. AI-DSS do not replace that judgment, they assist it. Where an AI-enabled capability is available, sufficiently reliable, and materially improves the implementation of precautionary obligations without imposing disproportionate operational costs, its use may become necessary to satisfy the requirement to take all feasible precautions. Conversely, where an AI system is unreliable or otherwise unsuitable for the operational circumstances, reliance on it may itself be inconsistent with the duty to take feasible precautions. The significance of AI under the principle of precaution therefore lies not in replacing human judgment but in expanding commanders’ ability to implement the law’s precautionary obligations. That conclusion, in turn, raises a further question. Should the humanitarian implications of military AI be assessed beyond the individual attack?
Evaluating AI at the Campaign Level
The preceding discussion argued that military AI should be evaluated according to whether it assists commanders in implementing precautionary obligations more effectively than would otherwise be possible under comparable operational conditions. Because that analysis focused on the implementation of precautions in individual attacks, it naturally raises a further question. If AI-DSS support military decision-making continuously throughout military operations, can their humanitarian implications be fully understood by examining individual attacks alone?
IHL assesses the legality of each attack individually, based on the information reasonably available to commanders in the circumstances ruling at the time. Accordingly, existing legal debates have focused predominantly on the humanitarian implications of AI in relation to individual targeting decisions. The question considered here is not whether that framework should change, but whether evaluating the humanitarian implications of military AI requires a broader perspective. AI may influence how military operations are conducted more broadly, with humanitarian consequences that cannot necessarily be captured through the assessment of individual attacks alone.
This broader influence reflects the role of military AI in supporting military decision-making throughout military operations rather than only at the moment an attack is authorized. As a result, while it may improve individual targeting decisions, the same capabilities may also increase operational tempo by enabling military objectives to be identified, verified, and engaged more quickly than would otherwise be possible. Consequently, AI may influence not only the quality of individual targeting decisions but also the overall conduct of military operations.
This gives rise to a potential humanitarian paradox. If AI may reduce the risk of civilian harm in each individual attack while simultaneously enabling commanders to conduct substantially more attacks over the same period, the aggregate humanitarian outcome becomes less straightforward. The reduction in civilian risk per attack may be offset by the increased number of attacks conducted, resulting in greater aggregate civilian harm despite improved performance in each individual engagement. Conversely, the same increase in operational effectiveness may enable military objectives to be achieved more rapidly, thereby shortening the duration of hostilities and reducing the overall humanitarian costs of the conflict. Shorter conflicts produce important humanitarian benefits by reducing civilian exposure to violence, displacement, and the disruption of essential services. While this is a possible outcome, some scholars argue that greater military efficiency may instead prolong conflicts, thereby undermining these potential humanitarian gains. Which of these competing effects ultimately predominates remains an open empirical question for which sufficient evidence is not yet available.
The purpose of identifying this operational dimension is not to suggest that either outcome is inevitable or to resolve this empirical question. Rather, it is to argue that assessing military AI exclusively through the legality and humanitarian consequences of individual attacks provides only a partial account of its humanitarian implications. This does not diminish the importance of assessing every attack in accordance with IHL. Instead, it suggests that this attack-level analysis should be complemented by considering how AI may influence the conduct of military operations across a campaign. Because AI is designed to support military decision-making continuously throughout military operations rather than only at the moment an attack is authorized, it may shape humanitarian outcomes beyond individual attacks. These broader operational effects therefore warrant careful legal and ethical consideration.
Conclusion
Recent debates on military AI have rightly focused on the humanitarian risks these technologies may create. Questions of accountability, meaningful human control, reliability, distinction, proportionality, and precaution will remain central as AI becomes increasingly integrated into military operations. Yet evaluating AI exclusively in terms of the risks it poses provides only a partial account of its humanitarian implications.
This post has argued that the principle of precaution provides the appropriate framework for a broader humanitarian assessment of military AI. These systems are designed to improve the collection, integration, and analysis of information relevant to military decision-making. They should therefore be evaluated not only according to the risks they pose but also according to their capacity to assist commanders in implementing existing IHL precautionary obligations. This perspective also requires a different standard of comparison. The relevant question is not whether AI performs better than an idealized human decision-maker, but whether AI-assisted decision-making enables commanders to implement precautionary obligations more effectively than unaided human decision-making under the same operational conditions.
Most importantly, the humanitarian implications of military AI should not be assessed solely through the legality of individual attacks. AI-DSS support military decision-making continuously throughout military operations, and their humanitarian implications may therefore extend beyond the cumulative assessment of individual attacks. Whether these operational effects ultimately increase or reduce the overall humanitarian costs of armed conflict remains an open empirical question. Recognizing that this question forms part of the humanitarian assessment of military AI is itself an important step toward a more complete legal and ethical framework.
This post neither assumes that AI will improve civilian protection nor suggests that existing legal rules require fundamental revision. It instead proposes a broader analytical framework that complements the prevailing emphasis on humanitarian risk by recognizing both IHL’s affirmative protective obligations and the broader operational implications of technologies that increasingly shape the conduct of contemporary warfare.
***
Oded Hen is a PhD candidate in the Department of Philosophy at Bar-Ilan University. His research focuses on moral philosophy, particularly the ethics of modern warfare and just war theory.
The views expressed are those of the author, and do not necessarily reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense.
Articles of War is a forum for professionals to share opinions and cultivate ideas. Articles of War does not screen articles to fit a particular editorial agenda, nor endorse or advocate material that is published. Authorship does not indicate affiliation with Articles of War, the Lieber Institute, or the United States Military Academy West Point.
Photo credit: U
