CommentaryOpenClipart-Vectors / Pixabay
AI-enabled weapons that select and engage targets on their own are already in the field, and international law has not settled who answers when they kill the wrong person.
Weapons that select and engage targets without a human pulling the trigger are no longer hypothetical, and the legal architecture governing who bears responsibility when they kill the wrong person has not caught up. US Department of Defense doctrine defines autonomous weapon systems (AWS) as systems that, “once activated, can select and engage targets without further intervention by an operator,” a category that now includes fielded hardware. In July 2026 the US Marine Corps selected the “Bullfrog” system, which detects and tracks aerial threats such as drones and automatically aims firearms at them. More broadly, the Maven Smart System (Maven), is being adopted by the US Department of Defense (DoD) to act as “the connective tissue between operators and sensor feeds, platforms … and algorithms [to improve] operational decision-making through human-machine teaming.” Maven can be integrated into any warfare processes, and is not limited to any specific hardware or software. Louis Mosley, head of UK and European operations at Palantir, which built the system, has described it as “instrumental” to US war efforts in Iran.
The Need for Human Supervision
The emergence of AI-enabled warfare has sparked debates about the extent to which we should retain human supervision over such combat. Some commentators take a more permissive view. For example, Tim McFarland of the University of Queensland Law School notes that “there do not appear to be any rules of [international humanitarian law] which necessarily preclude a degree of reliance on machines.” Moreover, he notes that “autonomous technologies endow weapon systems with some ability to detect information about their environment and respond [thus, accordingly] AWS may greatly reduce the need for … direct human intervention.” Shin-Shin Hua of the University of Cambridge is also sympathetic to the adoption of AWS, noting that they have become “a strong candidate for replacing human operators due to their ability to make decisions based on large amounts of data.” These positions, however, sit outside the prevailing weight of scholarly and diplomatic opinion.
In contrast to the above, the wider literature on this issue advocates for the retention of close human supervision over AI-enabled warfighting. Michael Boardman and Fiona Butcher of Dstl, the UK Ministry of Defence’s Defence Science and Technology Laboratory, call for the “design of systems with modes of operation that allow human intervention and require their input in specific steps of the targeting cycle.” This aligns more with the trend of international discourse, which is perhaps best encapsulated in the most recent United Nations (UN) Group of Governmental Experts (GGE) report on AWS. The GGE chair stated in the April 2026 report that “context-appropriate human judgement and control is needed to ensure the use and effects of [AWS] are in compliance with international law.” This conforms with the views of the UN secretary-general and the president of the International Committee of the Red Cross, who issued a joint statement based on “ensuring the ability for effective human supervision, and timely intervention and deactivation.”
Retaining human supervision is justified in part because AI systems encode and amplify the biases present in their training data and their designers. The State of Palestine raised gender bias in a 2023 GGE submission, warning that “a system may associate all military-aged males with combatancy … creating the risk that certain persons trigger an AWS strike although they are not legally permissible targets.” Similarly, there are dangers that AI incorporated into warfare processes will feature racial biases. This issue was raised by several other states in a GGE document stating that we “risk repeating, amplifying, or contributing to unjust biases [and that] … that human beings consciously or unconsciously encode their own biases into the programmes they write.” The idea that a person might be identified as a combatant partly as a result of the color of their skin, for example, is a worrying prospect. Thus, the retention of human supervision is critical.
If AWS must remain under human supervision, the harder question is what that supervision requires in practice. The GGE chair’s April 2026 summary offers the most concrete framework. The summary calls on states to: “ensure that [AWS] are operated under a responsible chain of command and control … ensure the effects of [AWS] can be adequately anticipated and controlled … ensure that mission parameters of [AWS] cannot significantly be modified by the system … ensure that [AWS] can be deactivated in a timely manner [and] limit the operation of [AWS] to a defined perimeter.” In short, for the GGE, appropriate human supervision involves a chain of command, anticipation of effects, fixed mission parameters, deactivation mechanisms and a clearly defined area of operation. If all these features are present, it should be possible to keep an AWS on a sufficiently tight leash during its use. Conversely, if one or more of these is lacking, then the use of an AWS could be said to be beyond appropriate levels of human supervision.
The GGE’s approach largely tracks the academic commentary. However, it lacks one of the points that has, arguably, been better captured in the literature: intervention. In other words, if all the criteria outlined by the UN GGE are satisfied, but to further guard against errors, to what extent should the human operator be able to intervene? On this point, Amanda Musco Eklund of Umeå University surveyed the views of some of the key contributors to the AWS debate and noted that the different options include: “the ability to cancel, suspend and modify attacks, overriding the system and the possibility of partial or complete deactivation [or] “sufficient time” for intervention, “potential for timely intervention” or “sufficient time for deliberation” [or] more absolute terms requiring the possibility to intervene “at any point in time” … or that the weapon must be under direct control and supervision “at all times.” Here, we can see a spectrum of the possible levels of “intervention” that could be worked into an AWS to ensure it remains under human supervision. This spectrum ranges from weaker to stricter, with “the ability to cancel, suspend and modify attacks” being less “interventionist” than the idea of direct control “at all times.” Given the weaknesses in AI, such as its gender and racial biases discussed above, a stricter approach seems preferable. This would mean human supervision throughout operation, coupled with the ability to intervene.
This piece has considered the rise of AI in warfare, specifically the phenomenon of autonomous weapon systems. Views differ, but the dominant position among states, international organizations, NGOs and commentators appears to be that human supervision over such systems must be retained. That position rests on a recognition of the weaknesses in AI, such as its tendency towards gender and racial biases based on those of its creators. To secure an appropriate level of human oversight, we must ensure that AWS: are grounded in a solid chain of command; are operated by people who can anticipate their effects; are given fixed mission parameters; are built with deactivation mechanisms; and operate within a clearly defined area of operation. In terms of the human operator, there is a spectrum around how “involved” they ought to be, however the safest option involves always placing an AWS under their direct control and subject to their intervention.
Elliot Winter is a lecturer (assistant professor) in international law at Newcastle University in the United Kingdom.
Opinions expressed in JURIST Commentary are the sole responsibility of the author and do not necessarily reflect the views of JURIST’s editors, staff, donors or the University of Pittsburgh.
