Toward a comprehensive understanding of civilian harm

As outlined above, the protection of civilians and civilian objects from the effects of hostilities lies at the centre of IHL principles and military doctrine related to distinction, precaution and proportionality.Footnote ¹⁷ Meeting these obligations throughout the JTC requires military decision-makers to understand not only what constitutes civilian harm but also its main causes.Footnote ¹⁸ In terms of legal elements of proportionality assessments, Van den Boogaard discusses the notion of “excessive” in his treatise on the subject.Footnote ¹⁹ More recently, Daniele and Sari have zoomed in on a debate about the meaning of “incidental”.Footnote ²⁰ Gillard is one of very few examples in the literature or military manuals examining the substance of the incidental side of the assessment.Footnote ²¹ Her paper places those discussions within a broader, less explored understanding of this aspect of assessment, one that encompasses “loss of civilian life, injury to civilians, damage to civilian objects, or a combination thereof”, commonly referred to as civilian harm. This approach has gained operational traction among several militaries in recent years. Drawing on documented cases, civil society organizations and scholars have also urged a broader, more nuanced interpretation of this clause that extends beyond immediate casualties and understands “injury to civilians” to encompass long-term, indirect and systemic effects.Footnote ²² Militaries, too, have begun to recognize the strategic, operational and tactical value of this wider approach, seeing it as integral to the legitimacy of their operations.Footnote ²³ These shifts are shaping doctrine on civilian harm mitigation and response (CHMR), a key element of civilian protection in armed conflict that applies in both asymmetric warfare, where civilians often live in or near conflict zones, and large-scale combat operations conducted in proximity to more dispersed civilian populations.Footnote ²⁴ CHMR involves actions to “prevent, deter, pre-empt, and respond to situations where civilians are targets of violence or under threat of violence”.Footnote ²⁵ Beyond being a legal, strategic and ethical imperative, CHMR strengthens operational legitimacy and can improve operational outcomes.Footnote ²⁶ Many military professionals now argue that embedding CHMR into operations enhances rather than hinders mission effectiveness, with some describing it as a “fourth strategic offset” in modern warfare.Footnote ²⁷

This article links CHMR approaches with the legal obligations of proportionality assessments in the JTC, offering an integrated perspective on both. Central to the CHMR shift, and adopted here, is a broader, more nuanced view of civilian harm and what this means in terms of legal understandings. Fully understanding civilian harm, both its forms and causes, is essential to complying with the duty of constant careFootnote ²⁸ and applying the rules of precautions in attack and proportionality in the JTC.

Historically, targeting processes have traditionally focused on analyzing effects on adversaries, with far fewer resources dedicated to assessing effects on civilians, civilian objects and the civilian environment.Footnote ²⁹ While militaries are beginning to acknowledge the policy value of better understanding the civilian environment, this article argues that targeting law demands this as well – for example, aligning ex ante civilian harm estimates (created from collateral damage estimation methods or a combination of intelligence sources) with ex post assessments, including battle damage assessments and after-action reports. This iterative process allows lessons from previous strikes to inform future decision-making. Harmonizing the metrics used in pre- and post-strike analysis so that the same elements of civilian harm are measured before and after an attack would significantly strengthen compliance with IHL. Indeed, the precautionary principle arguably requires such an approach.Footnote ³⁰ This perspective is set against the backdrop of increasing use of AI-DSS, where the extension of quantification logics increasingly shapes how civilian harm is understood and potentially programmed into certain AI-DSS and, therefore, how proportionality is evaluated and operationalized in targeting decisions.

The primacy of the duty of constant care and taking precautions in attack

Building on this discussion of civilian harm, the next question is how to avoid or in any case minimize it in armed conflict. This article adopts and extends a view, shared by certain scholars and practitioners, of assessing proportionality as a final legal safeguard, one that becomes relevant for consideration only after the duty of constant care has been complied with across the conduct of hostilities, all feasible precautions in attack have been taken, and some incidental harm remains unavoidable.Footnote ³¹ A complementary operational perspective confirms that only after precautions have been taken can it become clear what kind of civilian harm may occur, which can then be fed into any subsequent proportionality assessment.Footnote ³² To be clear, precautions must be continually reassessed as circumstances evolve (also within proportionality assessments), so proportionality as one of the core co-equal principles alongside distinction and precaution is never a singular, isolated step. Rather, in terms of the functionality of the proportionality rule outlined in Article 51(5)(b) of AP I, this article advances the argument that a proportionality assessment becomes relevant not as a first consideration in targeting but rather only once it is evident that, despite all feasible precautions having been taken, some incidental civilian harm is still expected to occur. Corn offers a useful visual illustrating how these concepts interact in targeting decisions, as shown in Figure 1.

Figure 1.

Targeting matrix, denoting the position of proportionality assessments at the bottom of the chart only after all feasible precautions have been taken. Source: Geoffrey S. Corn et al., The Law of Armed Conflict: An Operational Approach, 3rd ed., Wolters Kluwer, New York, 2006 (forthcoming).

As Figure 1 indicates, targeting is iterative and non-linear. Once targeting procedures begin, the law requires precautions to be taken first and often. However, certain interpretations of the law, shaped by the quantification logics embedded into military decision-making processes (MDMPs)Footnote ³³ over the past twenty-five years, have had the (perhaps unintentional) effect ofcentral evaluative standard, often at the expense of the separate and prior obligation to take precautions in attack as well as the overarching duty of constant care to spare civilians from the effects of conflict.Footnote ³⁴ As this author and Marta Bo have argued elsewhere, this risks contributing to a normative shift in interpreting legal obligations and the erosion of core protective mechanisms embedded in IHL and international law more broadly, diluting the emphasis on positive obligations to take measures to mitigate civilian harm in favour of manufacturing space for proactive justifications for its occurrence.Footnote ³⁵ By leveraging quantitative modalities and the increased speed and scale they provide, integrating AI-DSS into the JTC may further exacerbate this erosion.Footnote ³⁶

Although proportionality assessments have been described as Footnote equations ³⁷ or calculations,Footnote ³⁸ they are not mathematical problems to be solved.Footnote ³⁹ Within various phases of the JTC, outlined below, military decision-makers are required to undertake complicated analyses based on vast amounts of intelligence and information in order to assess and avoid or in any case minimize the potential for civilian harm before approving or effectuating strikes. In so doing, they must continuously qualitatively balance the military necessity of strikes against humanitarian concerns, take all feasible precautions to avoid or at least minimize civilian harm, and make ethical and legal judgements about what constitutes “excessive” harm in real time, often under intense pressure and time constraints.Footnote ⁴⁰ This article’s premise aligns with Wright’s conclusion that there is “no bright-line rule” for determining what constitutes excessive civilian harm in relation to anticipated military advantage.Footnote ⁴¹

Moreover, pure “objectivity” in these assessments is not required, possible, or necessarily desirable. The International Committee of the Red Cross (ICRC) Commentary on the Additional Protocols explains that the proportionality test is inherently subjective, granting commanders a “fairly broad margin of judgement”, and emphasizes that it should be guided primarily by “common sense and good faith”.Footnote ⁴² As Cohen and Zlotogorski observe, proportionality is so context-dependent that no uniform standard can be applied; each situation must be assessed individually.Footnote ⁴³ This means that, in practice, a military decision-maker must carry out a case-by-case, context-sensitive evaluation. Such assessments rely on legal reasoning, ethical judgement and situational awareness.Footnote ⁴⁴ Years of experience and levels of IHL training play a significant role in leading to reasonableness of decision-making, as does access to well-trained legal advisers.Footnote ⁴⁵

The standard of the “reasonable military commander” reflects the way in which the broader legal concept of reasonableness is enshrined in IHL.Footnote ⁴⁶ Henderson and Reece offer a deeper examination of the historical and judicial development of the term, according to which commanders, drawing on their experience and training, make proportionality judgements based on the information reasonably available to them at the time.Footnote ⁴⁷ However, this subjectivity is not without limits; it must align with what a hypothetical reasonable commander would have concluded under similar circumstances, grounding the assessment in both legal norms and practical military realities. This leads Henderson and Reece to conclude that the reasonable military commander standard reflects an “objective but qualified” approach.Footnote ⁴⁸ The standard acknowledges the complexities and uncertainties inherent in real-time decision-making during armed conflict, embedding a degree of subjectivity that reflects the commander’s training, situational awareness, experience, rationality, honesty and good-faith intent in interpreting ways to avoid or in an any case minimize harm to civilians while allowing for effective achievement of concrete and direct military advantage as a result of targeting decisions.Footnote ⁴⁹ Other factors, including checklists or auxiliary means of intelligence setting out requirements of information necessary to make decisions within the MDMP, can help to ensure a higher degree of objectivity in reaching the standard of reasonableness for military commanders, but they cannot be a substitute for context-appropriate human judgement and control.Footnote ⁵⁰

In other words, once it must be assessed, proportionality must be evaluated through a comprehensive harm lens – one that accounts not only for civilian fatalities but also for “injury to civilians”, including mental harm, displacement, loss of infrastructure and livelihood, and damage to cultural or environmental assets. These considerations are not reducible to mere data points,Footnote ⁵¹ and proportionality cannot be determined through fixed arithmetic or algorithmic values.Footnote ⁵² To operationalize this understanding, military training on the application of proportionality must emphasize analogical reasoning (encapsulated in the notion of the “reasonable military commander”) over rigid computational thresholds. Rather than quantifying civilian harm through abstract metrics, training modules should be built on comparison-based exercises drawn from past cases that emphasize the integrated approach outlined above: setting ex ante civilian harm expectations against actual ex post effects, informed by experiential knowledge and moral reasoning.Footnote ⁵³ All of these are areas in which human judgement excels. Some might argue that large language models (LLMs), with their strong pattern recognition abilities, could perform these tasks better than humans,Footnote ⁵⁴ but studies show that these models often miss subtle contextual meanings and that their performance declines further when they are compressed or made more efficient.Footnote ⁵⁵ This suggests that context-sensitive human judgement and control cannot technically be replaced by LLMs or AI-enabled systems. Moreover, even using them merely to assist humans carries risks, some of which are outlined below. Militaries working to integrate these systems within targeting operations should be aware of these risks and work to eliminate or minimize them; specifically in this instance, introducing LLMs to targeting cycles could create situations of over-reliance, leading to cognitive shifting, where human judgement deteriorates due to excessive trust in AI outputs. This too is discussed in more detail below.

While proportionality assessments are challenging, particularly for commanders under pressure, this difficulty reflects the high-stakes nature of the decisions involved. As Gillard observes, setting precise qualitative parameters may be difficult, but the determination is not impossible in practice.Footnote ⁵⁶ Proportionality must be understood as a case-by-case, context-specific balancing exercise, not a fixed metric reducible to algorithmic modelling.Footnote ⁵⁷ In the context of targeting operations, examined in the next section, this lens helps to reveal how such balancing is to be operationalized when commanders must weigh proportionality.

The JTC and proportionality assessments

Targeting procedures provide a structured methodology through which military forces identify, analyze and engage targets while complying with operational, legal and ethical obligations.Footnote ⁵⁸ While the precise formulations vary, many States and alliances use comparable targeting models. For example, both the United States and the North Atlantic Treaty Organization (NATO) employ a JTC consisting of six non-linear phases, and while the terminology and sequencing occasionally differ (NATO’s phase 1, for instance, explicitly incorporates the commander’s intent, objectives and targeting guidance), the underlying logic and functions of each step largely overlap. Likewise, other conceptual frameworks (such as “find, fix, track, target, engage, and assess”Footnote ⁵⁹) capture much of the same process. Because the US JTC is publicly articulated, doctrinally detailed and broadly mirrored in NATO doctrine and practice, this section uses it as a representative model to illuminate where and how key decisions, especially those related to proportionality assessments, are made within targeting procedures. That cycle generally includes the following six phases (see Figure 2):

1. 1. End-state and commander’s objectives: Defining strategic military goals and desired outcomes.

2. 2. Target development and prioritization: Identifying, verifying/validating and prioritizing targets based on intelligence and mission goals.

3. 3. Capabilities analysis: Assessing the available strike options and their effectiveness.

4. 4. Force assignment: Allocating specific military assets (e.g., air strikes, artillery, cyber operations) to engage the target.

5. 5. Mission execution: Carrying out the targeting operation while ensuring compliance with relevant laws and the rules of engagement.

6. 6. Assessment: Evaluating the effectiveness of the operation and adjusting for future operations if necessary.

Figure 2.

Joint Targeting Cycle. Source: US Department of Defense (DoD), Joint Targeting, Joint Publication 3-60, 28 September 2018, Fig. II-2, available at: www.esd.whs.mil/Portals/54/Documents/FOID/Reading%20Room/Joint_Staff/21-F-0520_JP_3-60_9-28-2018.pdf.

AI-DSS are reportedly increasingly being integrated and used at multiple phases of this JTC, particularly in target development and prioritization (phase 2), capabilities analysis (phase 3) and mission execution (phase 5).Footnote ⁶⁰ However, use of these systems raises concerns about whether human decision-makers can retain cognitive autonomy over targeting decisions when AI-DSS have been integrated, specifically within phase 4, or whether they will become overly reliant on algorithmic outputs because of the quantification logics embedded within and extended by these systems.

Vietnam and Cold War era: The “sci-tech” lens and the McNamara fallacy

The roots of quantification in warfare can be traced back to the Vietnam and Cold War era, a pivotal period in the rise of computational logics.Footnote ⁶⁴ Emerging data-processing technologies employed by the US military began to shape how complex social and military realities were perceived and acted upon. Advances in systems analysis, statistical modelling and computer-assisted decision-making converged with military strategy, embedding numerical metrics at the heart of operational planning. Schwarz describes this as viewing operations through the “scientific-technological lens”, a method that prioritizes and privileges the quantifiable aspects of war.Footnote ⁶⁵ This shift reflects a broader faith in data, where the perceived objectivity of quantification overshadows the complex qualitative dimensions of human conflict.Footnote ⁶⁶

Specifically, the Vietnam War marked a key shift for the United States toward quantification in military decision-making. The introduction of the Hamlet Evaluation System (HES), a data-driven tool measuring pacification progress through metrics like enemy presence and government control, is one example of this.Footnote ⁶⁷ Though statistically innovative, the HES was undermined by subjective inputs and data manipulation, failing to capture on-the-ground realities. This represented an example of what would become known as the McNamara or quantification fallacy, a critique of the over-reliance on quantifiable data in decision-making processes.Footnote ⁶⁸ The fallacy arises when decision-makers prioritize measurable variables while disregarding critical qualitative or human factors simply because they are not easily captured through measurable elements. At its core, the fallacy assumes that what cannot be quantifiably measured is irrelevant, which results in flawed reasoning. Particularly in the context of armed conflict, where the environment is highly complex and dynamic, this can produce grave operational, legal and ethical consequences, including increased risks of civilian harm.Footnote ⁶⁹

In this era, the United States also introduced the term “collateral damage” as a euphemism for civilian harm.Footnote ⁷⁰ As nuclear tensions rose, so did the use of collateral damage estimation methodologies (CDEMs), or analyses conducted by the military to estimate potential harm to civilians and damage to civilian property during an operation. Initially focused on infrastructure, CDEMs soon shifted toward estimating civilian deaths.Footnote ⁷¹ These estimates later formed the basis for non-combatant casualty cut-off values (NCVs), discussed below, which set quantitative thresholds that cognitively influence the interpretation of “acceptable” levels of civilian harm.

Network-centric warfare: Connecting everything to everything else

Building on the United States’ legacy of Vietnam-era quantification, and despite the limitations of those approaches, this evolution in military decision-making continued into the late twentieth century, culminating in the rise of network-centric warfare (NCW), which further embedded data-driven approaches at the heart of operational doctrine.Footnote ⁷² NCW aims to turn information superiority into combat power by linking sensors, decision-makers and shooters to enable shared awareness, faster decisions, higher operational tempo, and greater lethality, survivability and self-synchronization.Footnote ⁷³ NCW aims to redefine effectiveness through speed, precision and data integration, enabling commanders to act pre-emptively with optimized force configurations.Footnote ⁷⁴ Within this doctrine, the risk is that legal assessments undertaken by commanders will shift from qualitative, context-sensitive judgements to decision variables optimized by predictive modelling and surveillance data, where speed becomes a critical factor, as explored further in the discussion on AI-DSS below.

The “Global War on Terror”: Predictive analytics, CDEMs and NCVs

With NCW as the foundation, the post-9/11 shift in US military doctrine (mirrored by allied forces during coalition operations) toward counterterrorism and counter-insurgency, combined with rapid technological advances, accelerated the reliance on quantification logics and tools in targeting. The widespread use of armed drones, combined with algorithmic surveillance systems that assess threats and estimate civilian harm,Footnote ⁷⁵ paved the way for controversial “signature strikes”, targeting individuals based on behavioural “pattern-of-life” analyses rather than confirmed identities.Footnote ⁷⁶ These strikes reportedly relied on metadata and algorithms to infer threats, often without direct human verification, raising significant IHL concerns about the distinction between civilians and combatants and the obligation to take feasible precautions in attack.Footnote ⁷⁷ As Schwarz outlines, these technologies reshape human decision-making in war by embedding operators within a technological framework that projects enhanced, data-driven “superhuman” perception and an illusion of moral (and, as argued here, legal) certainty. However, this reliance on algorithmic profiling and quantification logic reduces complex legal and moral judgements to technical processes, enabling “signature strikes” where individuals may be deemed guilty until proven innocent, often only after death.Footnote ⁷⁸

The United States extended the CDEM highlighted above through a suite of analytical tools during this era.Footnote ⁷⁹ One of the most widely known, the Collateral Damage Estimation Tool, colloquially known as “Bugsplat”, uses imagery and data to estimate potential civilian harm.Footnote ⁸⁰ The Population Density Reference Table is a tool that projects likely civilian presence in target areas.Footnote ⁸¹ Other systems, such as the Digital Imagery Exploitation Engine and the Digital Precision Strike Suite Collateral Damage Estimation tool, employ algorithms to support strike planning by locating and characterizing targets, conducting choice of weapons and coordinate measurements, estimating collateral damage, and producing output graphics for databases.Footnote ⁸²

These models often omit key factors that are essential for a full understanding of civilian harm, such as the effects of secondary explosions and other reverberating effects.Footnote ⁸³ While these tools project an aura of precision and objectivity, critics contend that they risk reducing human lives to statistical abstractions, obscuring the moral gravity of lethal decisions.Footnote ⁸⁴ From a legal perspective, such omissions raise IHL concerns, particularly as regards the principles of distinction and proportionality, which require decision-makers to consider all foreseeable effects on civilians. Excessive reliance on algorithmic outputs, particularly in high-tempo decision-making environments, risks sidelining the qualitative and context-specific assessments required under IHL. The speed and efficiency introduced by AI-DSS can create pressure to act quickly, often at the expense of the careful, time-consuming evaluation needed to foresee, assess and minimize civilian harm. This tension between faster decision cycles and the legal obligation to account for all reasonably foreseeable harm can lead to systematic under-estimation of civilian risk.Footnote ⁸⁵ Accordingly, human judgement must remain central. Quantitative tools must aid, not replace, the deliberate weighing of civilian risk mandated by IHL. Decisions informed by AI-DSS should be complemented by other intelligence sources and contextual analysis, even when doing so slows the decision-making process. Preserving this deliberative space is essential to ensuring compliance with the principles of distinction, precaution and proportionality.

A final example of quantification logic is the use of NCVs, or policy thresholds that define the maximum number of civilian deaths that a commander is authorized to accept as “collateral damage” in a proposed strike before requiring higher-level approval. The United States introduced these metrics in 2011 to streamline proportionality assessments. NCVs sometimes set the threshold as low as zero, effectively prohibiting civilian casualties without senior authorization.Footnote ⁸⁶ During Operation Inherent Resolve, NCVs varied by target type and urban density – explicit thresholds of thirty were reported to be applied in operations like Iraqi Freedom and the targeting of Osama Bin Laden.Footnote ⁸⁷

As Corn explains, NCVs were not intended to determine what level of civilian harm is legally acceptable or proportional under IHL; rather, they were designed to establish cognitive and procedural thresholds that prompt commanders to escalate decisions.Footnote ⁸⁸ Yet simply setting numerical limits shapes how proportionality is assessed, effectively framing when anticipated civilian harm warrants heightened scrutiny relative to expected military advantage.

NCVs function as policy tools that may have the effect of overriding nuanced deliberation, reducing complex legal and ethical reasoning to mathematical or algorithmic calculations. They effectively codify an acceptable level of civilian deaths and risk anchoring that standard into any legal analysis of incidental harm.Footnote ⁸⁹ By framing anticipated civilian harm narrowly, as a quantifiable, acceptable cost outweighed by military advantage, NCVs risk distorting core IHL principles, reflecting a fundamental inversion of the primacy of the duty of constant care and precautions in attack and subordinating these legal requirements to proportionality logic. This approach risks shifting civilian protection from a paramount obligation to a negotiable or even optional factor, normalizing the perception that civilian harm is routine and unavoidable.

By introducing a fixed numerical threshold for acceptable civilian casualties, the metric also imposes a veneer of objectivity on a judgement that is inherently context-sensitive and qualitative.Footnote ⁹⁰ Although the intention is for this quantitative benchmark to guide decisions as they scale up through the chain of command, the anchoring effect of a numerical value (discussed in more depth below) exerts significant cognitive influence within MDMPs. The broader integration of AI-DSS risks accelerating this trend, which could further undermine efforts to prevent or minimize civilian harm. Though details are scarce, the Pentagon had officially abandoned NCVs in its doctrine by 2018, citing their ineffectiveness; however, their legacy continues to shape allied practices and military operations today.

The modern era: AI-DSS, quantification and the JTC

Many militaries worldwide are developing AI-DSS and, in some cases, deploying them in active conflicts.Footnote ⁹¹ These systems have evolved beyond basic computational tools into highly advanced technologies that collect and analyze vast amounts of battlefield data, generate predictive models, and provide information that can assist in making targeting decisions. Their capabilities include data synthesis, as they can rapidly sift through satellite imagery, drone footage, intelligence reports and signals intelligence to assist in demarcating patterns that help militaries to evaluate battlefield conditions and increase situational awareness.Footnote ⁹² Machine-learning algorithms that can be fed into AI-DSS can be programmed to identify movement patterns, target profiles or “threats” at greater speed, and therefore scale, than human analysts. And finally, in some scenarios, AI models can simulate potential battle scenarios, feed into CDEMs and offer suggestions for optimization of strike strategies.Footnote ⁹³ AI-DSS can analyze vast intelligence inputs, detect patterns and generate an overview of potential risks,Footnote ⁹⁴ using predictive models to estimate enemy behaviour, civilian harm and operational outcomes based on historical data.Footnote ⁹⁵ Despite all these purported benefits, however, these systems also introduce a host of new complexities and risks, as discussed in the following section.

Speed and scale: Risks of reinterpreting IHL through AI-DSS

Fighting at machine speed introduces risks and raises important questions about the trade-offs involved. Trying to embed ethical and legal judgements, or portions thereof, into formulas and algorithms (arguably not even possible from a technical engineering perspective),Footnote ¹⁰¹ risks stripping away the qualitative context needed for conducting proportionality assessments. As Pratzner observes, while lower-level functions might be able to partially be automated, core tasks such as target vetting, validation and target nomination – all forming part of a reasonable commander’s judgement – demand time and deliberation.Footnote ¹⁰² More generally, there are concerns about the role of these systems and the structure of human–machine teaming within the JTC regarding human cognitive agency over the MDMP.Footnote ¹⁰³ Specifically, and most relevant to this article, the use of AI-DSS raises critical concerns about how these systems may be reshaping the cognitive processes through which commanders make decisions.Footnote ¹⁰⁴

Unlike the full spectrum of complex human reasoning, which includes deductive and inductive thinking, decision-making, and problem-solving,Footnote ¹⁰⁵ AI-DSS rely on more constrained computational approaches. Machine-learning models, including transformer-based neural networks designed to process large volumes of sequential data such as text (e.g., LLMs), operate within these narrower, task-specific parameters. They can process entire sequences in parallel, enabling greater efficiency and speed, but they function through probabilistic modelling and operate within predefined parameters set by their design and intended purpose, even if shaped by human input.Footnote ¹⁰⁶ Because these systems are inherently limited in their ability to engage with the full range of contextual, legal, moral and experiential factors necessary for “a thorough assessment in good faith of all the different components of [a given] rule as well as the circumstances at the time”,Footnote ¹⁰⁷ algorithmically derived recommendations can contribute to a false veneer of objectivity and precision, reinforcing the illusion of quantifiable and data-driven accuracy in targeting decisions. This illusion is amplified by the speed and scale at which such systems are able to operate, forcing humans to keep pace and increasing the risk of automation bias and over-reliance on system outputs in the pursuit of technologically enabled “clean” warfare – which, as Assaad and Williams highlight, is a fundamentally unattainable aspiration.Footnote ¹⁰⁸

Over-reliance on AI-DSS may also encourage commanders to progressively disengage from the exercise of critical judgement, leading to cognitive offloading of their tasks and ultimately deskilling.Footnote ¹⁰⁹ This disengagement will be reinforced as decision-makers increasingly defer to AI outputs, trusting the system’s apparent rigour over their own nuanced judgement. Over time, this can weaken their ability to critically evaluate contextual factors, interpret ambiguous information or recognize subtle cues that a computational approach cannot capture.Footnote ¹¹⁰ The result is a risk of gradual erosion of domain-specific expertise, leaving human operators less capable of intervening effectively when algorithmic recommendations conflict with legal, ethical or operational considerations.

Rather than drawing upon their own training, operational experience, contextual understanding and sensitivity to legal and moral nuance, military decision-makers risk functioning as mere endorsers of machine-generated outputs so complex that they are impossible to trace or understand, thereby diminishing the role of human agency and responsibility in decisions.Footnote ¹¹¹ This carries profound ethical and legal consequences and ultimately risks rendering the commander unable to act based on their own decisions and in accordance with IHL, particularly in conducting a lawful proportionality assessment.

When programmed into AI-DSS that assist in calculating proportionality, such a numerical approach encouraged by quantification logics risks becoming self-reinforcing, as the technology can lend a false sense of precision and a veneer of legality to decisions that, in reality, demand qualitative human judgement. This process effectively shifts decision-making away from human commanders, through their own cognitive faculties, and towards a reliance on system-generated outputs, fostering a culture in which suggestions by the AI-DSS risk being “rubber-stamped”Footnote ¹¹² by the human rather than contextually qualitatively attained or critically interrogated.

While commanders have long relied on tools such as CDEMs to support proportionality assessments, these tools can only ever provide a partial view. As experts emphasize, interpreting CDEM outputs depends on practised judgement, intuition, and sound legal and ethical reasoning.Footnote ¹¹³ However, the growing use of AI-DSS in the JTC due to aspirations for an ever-increasing tempo in the MDMP risks shifting this balance away from human deliberation and towards increasingly data-driven processes. While AI-DSS can assist human decision-makers by providing rapid data analysis, probabilistic predictions and pattern recognition, they remain inherently limited in capturing the context-specific, legal and ethical nuances that are critical to proportionality assessments. As such, they can at best support, but never replace, the exercise of informed human judgement in targeting decisionsFootnote ¹¹⁴ – and it is essential to be aware that even as assistive tools, these systems introduce risks, including automation bias, over-reliance on algorithmic outputs, and cognitive offloading, which can progressively erode operators’ critical reasoning and domain expertise.

Systemic risks that AI-DSS may pose to IHL interpretation in targeting

The ICRC’s 2024 report on International Humanitarian Law and the Challenges of Contemporary Armed Conflicts highlights that the erosion of protective elements of IHL, coupled with the introduction of new technologies, creates systemic risks for civilians.Footnote ¹¹⁵ Tools that rely on preset numerical thresholds or predictive algorithms can reinforce these risks by prioritizing speed, scale and computational efficiency over context-sensitive, case-by-case legal judgement. When proportionality and distinction are reduced to calculable metrics, even in the absence of case-specific examples, there is a heightened danger that civilians may be exposed to harm which could otherwise have been mitigated through nuanced human assessment.

Operationally, such practices risk recalibrating the application of IHL principles. When high levels of incidental harm are pre-approved through elevated NCVs, especially for lower-level targets, the primacy of the duty of constant care and the precautionary obligation to avoid or minimize civilian harm become displaced. Because targeting is an iterative process, decisions made in phase 4 of the JTC influence earlier steps in subsequent cycles, including CDEM modelling in phase 3. As a result, “acceptable” levels of harm may progressively shift upward, normalizing greater civilian harm and lowering the evidentiary threshold for target verification. This would disregard or even circumvent the duty of constant care under Article 57 of AP I, sidelining precautionary obligations in favour of proportionality assessments and thereby recasting the MDMP itself.

The integration of quantification logics into AI-DSS illustrates the socio-technical reality that operational tools shape how legal and ethical norms are engaged in practice.Footnote ¹¹⁶ Efficiency, speed and data optimization risk outweighing humanitarian considerations, suppressing critical scrutiny of algorithmically generated outputs. Rather than providing a genuine legal justification, reliance on numerical thresholds can create the illusion that decisions are legally or ethically grounded simply because they are algorithmically derived. This reliance risks sidelining or obscuring the careful, context-sensitive legal assessments that IHL requires, leaving human operators dependent on computational outputs while potentially neglecting critical legal considerations. As numerical thresholds become routine, they institutionalize the notion that civilian harm is inevitable and manageable through algorithmic means, further entrenching this reliance.Footnote ¹¹⁷

As AI-DSS systems shape and constrain the informational environment in which commanders operate, the subjective dimensions of proportionality assessments risk being sidelined.Footnote ¹¹⁸ In terms of shaping and constraining, AI-DSS may present a commander with a limited set of recommended courses of action or nominated targets without revealing how many alternatives exist, how the system generated or ranked them, or whether they are truly optimal. In this sense, decision-making can resemble “looking at the world through a straw”, where the commander’s perspective and options are constrained by the system’s narrow and opaque outputs. Incidental civilian harm is codified into numerical tolerances, via NCVs, predictive algorithms or both, reflecting a broader “datafication” of warfare that threatens to erode IHL’s protective aims.Footnote ¹¹⁹ In this setting, human commanders shift from being central legal decision-makers to mere overseers of algorithmic processes. Preserving the integrity of proportionality as a legal standard requires critically examining how AI-DSS and related practices are reshaping the balance between human judgement and machine-driven analysis in armed conflict. Unlike AI-DSS parameters, human cognition allows for situational flexibility, adapting to unforeseen circumstances, interpreting ambiguity and applying legal reasoning in complex scenarios.Footnote ¹²⁰ Conversely, unlike human cognition, AI systems are ill-equipped to grapple with moral and ethical deliberation, interpret ambiguity, or assess adversary intent and strategic objectives.

From a technical standpoint, algorithmically guided systems will always lack the flexibility to engage with the qualitative dimensions of military decision-making. Their reliance on deterministic data processing limits their ability to navigate uncertainty, deception or incomplete intelligence, factors that are often central to operational judgement.Footnote ¹²¹ As Greipl notes, AI’s “forecasting” is based on data analytics of past behaviours and lacks context and human logic.Footnote ¹²² This renders AI systems unable to properly scrutinize, as required by IHL, whether a proposed objective is a lawful target or whether anticipated incidental harm will outweigh the expected military advantage under the circumstances prevailing at the time. While the predominant worry outlined in this article relates to how AI-DSS may skew proportionality analyses, their potential to undermine distinction is equally significant. Given that misidentification is a primary driver of civilian harm, designers, developers and deployers need a clear understanding of these systems’ error tendencies and must adopt measures to avoid, or at minimum mitigate, such misrecognition, as this understanding also feeds into proportionality assessments, should they become relevant.Footnote ¹²³

AI-DSS are inherently constrained by the scope and quality of their training data.Footnote ¹²⁴ These limitations mean they are poorly equipped to navigate environments marked by uncertainty, incomplete intelligence or ethically complex dilemmas, all three of which are ubiquitous within the context of armed conflict. In contrast, military commanders are assumed to have the training, capacity and legal obligation to interpret ambiguous information and exercise legal and moral judgement in nuanced and context-dependent situations.

Cognitive impact of AI-DSS: Automation bias, anchoring, offloading and deskilling

Gunneflo and Noll argue that historically, proportionality reasoning has functioned as a mechanism through which new technologies are lawfully integrated, largely through human interpretive judgement. Past innovations left the legal decision-making process intact, with humans still applying existing law. Digital decision support in the military, however, marks a different kind of shift: it enters the cognitive process of legal reasoning itself, potentially displacing the human role to the margins and creating a more direct link between law and technology. AI-DSS, Gunneflo and Noll contend, represent a technological leap that works from within, modifying or even displacing human judgement in unprecedented ways.Footnote ¹²⁵ Woodcock importantly highlights that while human decision-making is not free from bias, the introduction of AI-DSS brings new forms of bias and alters the very structure of decision-making.Footnote ¹²⁶ This shift reflects a broader cognitive transformation in warfare, one that risks marginalizing human judgement and displacing it from its central role in lawful and ethical military operations.

Concretely, the integration of AI-DSS into the JTC can introduce biases that can affect a commander’s ability to conduct legally sound proportionality assessments. These systems can subtly transfer decision-making authority from the performance of human judgement to the acceptance of algorithmic output, with knock-on effects for accountability and legal responsibility. The risk of (partial) cognitive offloading – or transferring reasoning tasks to external systems and reducing the mental effort required for decision-making – arises from designs aimed at efficiency and reducing “friction points”, as Woodcock describes.Footnote ¹²⁷ Because AI-DSS can rapidly process large volumes of data and generate recommendations, commanders may be tempted to assume that the system has accounted for all necessary variables, creating a feedback loop of growing dependence on AI tools. This could raise questions of legal compliance and create operational and strategic risks including erosion of human oversight, accountability gaps, operational dependency and heightened legal exposure.Footnote ¹²⁸ Mitigating these risks requires sustained training in, and consideration of, IHL compliance at every stage of system design, development, deployment and decommissioning, advancing a model of responsibility or legality by design.Footnote ¹²⁹

Potential biases introduced by AI-DSS

One way in which cognitive offloading manifests in proportionality assessments during the JTC is through automation bias – that is, the tendency to place undue trust in AI recommendations without critical scrutiny.Footnote ¹³⁰ This bias can lead to legally non-compliant decisions, especially when AI systems produce errors or uncertainties such as miscalculating risk, omitting qualitative factors or misidentifying a target.Footnote ¹³¹ In the JTC, a risk arises when vast targeting lists are generated: the speed of operations can create time pressure that leads to skipping target validation or verification altogether.Footnote ¹³² Treating AI-DSS outputs as optimized solutions can erode commanders’ critical engagement and heighten automation bias, especially in the later JTC phases (4 and 5) where precautions and proportionality must be (re)assessed under evolving conditions.Footnote ¹³³ Simply put, at machine speed, the space for moral and contextual legal reasoning shrinks, creating the risk that human oversight will become little more than a procedural rubber stamp.

Anchoring bias is another cognitive shift that AI-DSS can introduce, directly affecting proportionality assessments. Anchoring occurs when an initial piece of information disproportionately influences subsequent decision-making. In target selection and nomination, for example, an algorithm might place a certain individual at the top of a targeting list, anchoring perceptions of that target’s military value.Footnote ¹³⁴ Similarly, casualty or damage estimates from CDEMs in phase 3 may lead commanders in phase 4 to subconsciously adjust their judgements around this initial figure (even when new information suggests a different conclusion) despite their legal obligations to take all feasible precautions in attack in order to avoid or in any event minimize civilian harm.

NCVs are a concrete example of anchoring bias in practice. Once commanders internalize a set threshold of “acceptable” civilian harm for higher approval, it can shift what is perceived to be legally necessary or feasible under the duties of precaution and constant care. When proportionality becomes a numerical exercise, where figures like twenty, thirty or even hundreds of civilian deaths are weighed against eliminating a single target, it raises fundamental questions about whether the principles of precautions in attack and proportionality are truly being upheld, and whether such an approach is compatible with IHL’s object and purpose of limiting the effects of armed conflict by protecting civilians and restricting means and methods of warfare.

A final cognitive concern is deskilling or (meta)cognitive erosion, which is the gradual loss of a commander’s ability to conduct complex assessments when repeatedly following AI recommendations without exercising independent judgement. Empirical evidence has shown that increased reliance on automated support erodes individuals’ knowledge and undermines their confidence in making independent decisions.Footnote ¹³⁵ For example, deskilling through over-reliance on AI has recently come to the attention of the medical field, with one study showing that doctors’ ability to detect cancer devolved through AI-enabled decision support.Footnote ¹³⁶ This risk is not limited to the medical field and will be present across all sectors where these technologies are integrated.Footnote ¹³⁷ Translated to the military context, over time, this raises serious questions about long-term operational readiness and the capacity to make critical decisions without AI assistance.Footnote ¹³⁸

Returning to the earlier discussion on aligning CDEMs and battle damage assessment and after-action reports for a fuller understanding of civilian harm, in practical terms, during phase 6 of the JTC, the duties of constant care and of taking precautions in future attacks require comparing anticipated harm with actual effects and feeding that information into subsequent proportionality assessments. If commanders lose this evaluative capacity through over-reliance on algorithmic tools, whether by failing to incorporate a comprehensive understanding of civilian harm into battle damage assessments or after-action reporting or by skipping or condensing these steps under operational time pressure, cognitive offloading and deskilling can occur. This not only risks non-compliance with the proportionality rule but also undermines compliance with rules related to the precautionary principle and the duty of constant care.Footnote ¹³⁹