In response to the Hamas-led attack on Israel on October 7, 2023, Israel initiated the military operation ‘Iron Swords’ in the Gaza strip. Significant to the operation was the extensive use of artificial intelligence (AI), particularly through the AI-based decision-support systems (AI-DSS) such as Lavender, Where’s Daddy? and The Gospel.
These AI systems were reportedly used in targeting and surveillance processes, assisting military decision-makers in determining who and what to target, as well as when and where.
The integration of AI-DSS in military conflict has raised fundamental questions concerning their impact on military decision-making processes: in particular, questions related to how human judgement is exercised in such processes, and how military personnel can ensure the protection of civilians, which lies at the heart of International Humanitarian Law (IHL).
While IHL establishes the legal framework governing the conduct of hostilities, Just War Theory (JWT) provides an additional normative framework for examining moral judgement and responsibility in military decision-making. In particular, its jus in bello principles (also referred to as just conduct) of discrimination and proportionality offer ethical assessments for the protection of civilians and on the use of force during armed conflict. Responsibility is thus framed as a matter of moral judgement rather than as something determined solely by legal standards.
This blog examines how the operational use of AI-DSS in modern warfare may affect human moral judgement and responsibility in military decision-making. Using JWT as a normative framework and drawing on examples from the early stages of the Israel–Hamas conflict (2023–2024), it explores how the use of AI-DSS may challenge the ethical assessments underlying the principles of jus in bello. To understand how these challenges emerge in human–machine interactions, the concept of distributed responsibility is applied as an analytical lens.
Just War Theory and the Challenge of Distributed Responsibility
Proportionality and Discrimination
Within the Just War Tradition, ethical assessments of jus in bello are addressed through the principles of proportionality and discrimination. Proportionality concerns whether the expected costs of an attack—primarily the anticipated harm inflicted upon civilians and civilian infrastructure—are outweighed by the anticipated military advantage. Discrimination, also referred to as distinction, concerns the obligation to distinguish between civilians and combatants during armed conflict. Common to both principles is that they require individual and context-sensitive assessments and depend upon military personnel exercising human moral judgement prior to military action.
Today, AI systems may increasingly contribute to military decision-making processes in which such ethical assessments must be made, by generating recommendations that are subsequently evaluated by military decision-makers. As a result, decision-making processes are increasingly characterised by interactions between humans and AI systems, commonly referred to as human–machine interaction.
Distributed Responsibility
Relevant to examining human–machine interaction is the notion of distributed responsibility, which applies to the interaction of multiple actors rather than the actions of a single actor.
In military decision-making, this becomes particularly relevant as AI-DSS play an increasing role in shaping the informational basis upon which human decisions are made. Importantly, this distribution does not necessarily remove human responsibility but reflects how different human actors and systems contribute to the processes through which decisions are made. Distributed responsibility therefore provides an analytical lens for examining how human–machine interaction may reshape the conditions under which qualitative and context-sensitive moral judgement, required by proportionality and discrimination, is exercised.
AI-DSS in Practice: The Israel-Hamas Conflict (2023–2024)
During the early stages of the conflict, the Israel Defence Forces (IDF) reportedly employed at least three AI-DSS: Lavender, Where’s Daddy?, and The Gospel. Each supported military decision-making at different stages of the targeting process.
Lavender was reportedly employed to identify Palestinians with presumed affiliations to Hamas or Palestinian Islamic Jihad (PIJ), assigning individuals a score from 1 to 100 based on their estimated likelihood of affiliation. During the first weeks of the conflict, the system reportedly contributed to marking as many as 37,000 Palestinians as potential targets. Sources further report that military personnel merely functioned as “rubber stamps” and devoted as little as 20 seconds to reviewing individual targets—mainly to determine whether the target was a male—despite a known error rate of approximately 10%. Additionally, former intelligence officers stated that Lavender’s outputs were to be treated as “if it were a human decision”.
Where’s Daddy? complemented Lavender in the targeting process by tracking the geographical locations of individuals identified as potential targets. Using mobile phone location data, the system reportedly notified military decision-makers when marked individuals entered their private residences, where attacks could subsequently be carried out. Such strikes reportedly included the use of unguided (‘dumb’) bombs, increasing the potential for civilian harm in the surrounding area.
Gospel, by contrast, was employed to identify infrastructure as potential targets allegedly used by Hamas and/or PIJ. In addition to military targets and family homes, these included so-called power targets. That is, civilian structures whose destruction was intended to exert indirect pressure on Hamas through the civilian population. The use of power targets further reflects a pre-existing targeting logic associated with the so-called Dahiya Doctrine. The doctrine has been associated with the use of disproportionate force against infrastructure in areas where armed groups operate.
Against this backdrop, a former intelligence officer described Gospel as a “mass assassination factory”, reflecting both the speed and scale of target generation and the system’s role in automating a pre-existing military logic.
Identifying Distributed Responsibility Within the Use of AI-DSS
The integration of AI-DSS into military decision-making may result in the distribution of different elements of the decision-making process across humans and AI systems, making it increasingly difficult to determine where responsibility for a decision resides.
With Lavender, the generation of targets occurred through algorithmic predictions about an individual’s presumed affiliation with an armed group, based on numeric scores, rather than through human-based analysis alone. As military personnel were reportedly not required to thoroughly verify or examine Lavender’s underlying intelligence data, the target selection process and the rationale behind military decisions appear to have been substantially shaped by the system rather than by human decision-makers alone. Moreover, accounts of military personnel functioning as a “rubber stamp” and treating AI outputs as equivalent to human intelligence raise questions about the extent of independent human judgement.
This distribution becomes even more apparent when Lavender is considered in addition to Where’s Daddy?. While Lavender contributed to determining who could be a potential target, Where’s Daddy? subsequently contributed to determining where and when that individual could be targeted.
In this light, the use of AI-DSS in the Israel-Hamas conflict (2023-2024) highlights how a network of multiple AI-DSS used together may contribute to undermining human judgement in decision-making. Although the decision to authorise a strike may formally have remained human, the AI systems may have played a significant role in shaping the underlying basis on which that decision was made.
From Distributed Responsibility to Ethical Judgement
The distribution identified above may therefore have implications for the exercise of human moral judgement required by the principles of proportionality and discrimination.
Lavender illustrates this tension. Reports of military personnel devoting as little as 20 seconds to reviewing targets, despite a known error rate of approximately 10%, suggest limited opportunities for the individual and context-sensitive assessments required by the principle of discrimination. Rather than eliminating human judgement, reliance on Lavender may therefore have constrained the conditions under which it could meaningfully be exercised by increasing the speed and pressure of targeting decision-making.
This becomes particularly relevant when Lavender is considered alongside Where’s Daddy?. While Lavender contributed to identifying targets, Where’s Daddy? contributed to determining where and when they could be targeted. Reports that dumb bombs were subsequently used against some targets further raise questions of proportionality, as weighing anticipated military advantage against expected civilian harm ultimately remained a human judgement.
Gospel raises a related concern, as its generation of power targets may have contributed to operationalising and scaling a pre-existing military targeting logic associated with the use of disproportionate force.
Viewed through JWT, these examples therefore raise questions not only about the outcomes of military decisions, but also about how the ethical assessments required by proportionality and discrimination are exercised throughout the decision-making process.
Across the uses of these systems, we see a pattern: AI-DSS did not replace human decision-makers, but their use appeared to have shaped significant parts of the informational basis upon which humans made decisions, including on proportionality and discrimination. From a JWT perspective, the central concern is not simply whether humans remain involved, but whether the integration of AI-DSS in military decision-making processes challenged military personnel’s capacity to exercise the context-sensitive moral judgement required by the principles of proportionality and discrimination.
Therefore, ensuring human involvement does not neccesarily ensure human moral judgement. Developments in the use of AI-DSS point towards a broader challenge in modern warfare: the challenge of safeguarding human capacity for qualitative and context-sensitive judgement required by the ethical principles of jus in bello governing just conduct in war.
To address this development, recent debates point to the need for new approaches that address both the technical aspects of AI-DSS and the conditions necessary to strengthen human decision-makers’ ability to critically engage with the outputs from the systems. This includes ensuring that human operators are able to understand and question system outputs, cross-check them against other sources, and retain sufficient time to exercise meaningful human judgement.
About the author
Alexander Hedegaard is a student assistant at the Center for War Studies and the Department of Political Science at the University of Southern Denmark.
Featured image credit: Cristóbal Ascencio & Archival Images of AI + AIxDESIGN / Better Images of AI / CC BY 4.0