1. IntroductionFootnote 1
The morality, legality and prudence of integrating artificial intelligence (AI) into war are the subject of passionate, ongoing debate. In international forums, this debate has centered primarily on autonomous weapons (sometimes termed “killer robots”) – systems that, upon activation, can select and engage targets without further intervention by a human operator. Recent conflicts in Ukraine and Gaza have expanded the military AI debate to include battlespace management (Cronin, Reference Cronin2023) and targeting systems (Schmitt, Reference Schmitt2024).
AI in decision-making over the resort to force has yet to receive the same level of scrutiny.Footnote 2 In this article, I consider this potential application of AI from a Just War perspective. I evaluate two principles from this tradition: (1) the jus ad bellum principle of “reasonable prospect of success” and (2) the more recent jus ad vim principle of “the probability of escalation.” The former holds that war should not be waged when the likelihood of victory is remote, given the ruinous costs of violent conflict. The latter governs force-short-of-war, discouraging “limited” strikes when they are likely to escalate into open war. More than any other principles of Just War, these prudential standards seem amenable to the probabilistic reasoning of AI-driven systems.Footnote 3 Epistemic gaps can be bridged to allow for more reasoned analysis of the merits, cost and escalatory potential of force, and destructive human impulses tamed to minimize the likelihood of unjust violence.
I argue that optimism in the potential of AI-optimized decision-making is largely misplaced. We need to cultivate a tragic sensibility in our thinking and practice of war – a recognition of the inescapable limits of foresight, the permanence of uncertainty and the dangers of unconstrained ambition.Footnote 4 This sensibility is especially important when navigating the future-oriented questions of “reasonable prospect of success” and “probability of escalation.” We must be cautious about AI and machine-learning techniques precisely because of their intuitive appeal in this context. Exaggerated and undue confidence in the efficacy of these systems will blind us to their technical limits. It will also, more seriously, obscure the deleterious impact of AI on the process of resort-to-force decision-making; its potential to suffocate the moral and political wisdom so essential to the responsible exercise of violence on the international stage.
Before I unpack these arguments, a qualification is needed. Just War has always been, and remains, a contested tradition. Virtually every element of commencing, fighting and ending war, including the two I consider here, is debated between and within rival Just War accounts. My aim here is not to settle, or even really grapple with, these debates. Instead, I focus on an antecedent question – what dispositional qualities are needed to effectively and responsibly navigate ethical questions over the resort to force? My argument is that AI is likely to degrade these qualities.
2. AI and the reasonable prospect of success
The Just War Tradition has historically occupied the moral middle ground between pacifism and war-without-limits.Footnote 5 In contrast to pacifism, Just War asserts that states and certain non-state actors may have the right, and on occasion, even the moral duty, to engage in armed conflict. In contrast to war-without-limits, however, Just War conditions this right, articulating a set of moral standards to be followed in the commencement (jus ad bellum), conduct (jus in bello) and termination (jus post bellum) of war. War can only ever be a bad outcome. Just War recognizes this, but also acknowledges the existence of worse outcomes, the alleviation of which may, in rare instances, be worth a war. The battlefield can only ever be cruel. Just War recognizes this also, but works to reduce such cruelty as much as feasible. It is an imperfect and morally unsatisfying framework, yet a critically important one for those committed to lessening both the frequency and destructiveness of armed conflict.
To measure the likely impact of AI on resort-to-force decision-making and Just War compliance, I first consider the jus ad bellum principle of “reasonable prospect of success.” This principle stems from a recognition of the enormity of war. Armed conflict, even when waged for just reasons by just parties, ruins the lives of the innocent at scale. The harm produced is routinely of such magnitude that state and non-state actors should abstain from its commencement when success is unlikely. The alternative, the purposeless unleashing of war’s destructiveness, is “not merely imprudent but also unethical” (Eckert, Reference Eckert, Gentry and Eckert2014, p. 63). The principle of “reasonable prospect of success” speaks to the tragic qualities of international politics. Not all injustices can be remedied through force, or at all. Sometimes the most moral course is for the weak (and the strong) to suffer what they must.
Just War, to repeat, is more contested tradition than settled theory. So-called traditionalists and revisionists disagree over a range of critical issues: whether or not war is a “special moral sphere” (Frowe, Reference Frowe2014, p. 123), distinct from normal society and morality; what, if any, war rights are retained by the “unjust”; and whether formulations of good conduct are best determined via abstract hypotheticals or practical experience and history.Footnote 6 This debate extends to the principle of “reasonable prospect of success.” Since its integration into the Just War Tradition in the Sixteenth Century (Howard, Reference Howard and Howard1979, p. 8), the principle has faced significant criticism, in relation to both its parameters (how should one define success) and premise (are some struggles worth waging even when hopeless).Footnote 7 While much of this criticism is valid, the principle remains an important one, pushing belligerents to reflect on the tragic fact that “virtuous intention, unmoored from prudence, can sometimes be as destructive, or even more destructive, than outright malice” (Pelopidas & Renic, Reference Pelopidas and Renic2024, p. 211).
Could we harness AI to better evaluate our prospects of military success and, in doing so, minimize the gamble of war, or is this an illusory hope? I argue the latter. As I outline in the remainder of this section, AI-driven systems are more likely to distort than enhance our understanding of the viability and meaning of success.
2.1. Broken answers
To fairly assess the value and limits of AI in this context, we should first consider the dismal historical record of human agents. For as long as war has been fought, belligerents have overestimated the probability of victory and underestimated the cost. This not only applies to unjust wars of aggression, but also ethically motivated humanitarian intervention and self-defense. History is replete with examples of defensive wars of folly – pyrrhic victories, bloody stalemates or outright military defeats favored over more prudent alternatives, such as non-violent resistance or diplomatic compromise. Ned Dobos explores this over-confidence in his book, Ethics, Security, and the War Machine, highlighting a range of cognitive, psychological and moral factors that too often lead human decision-makers to overestimate the speed and ease of military success (Dobos, Reference Dobos2020, pp. 81–84).
Human analysts have also been guilty of the inverse – overestimating the competency of an unjust aggressor and wrongly pushing the just to preemptively sue for peace. The Russo-Ukrainian war is but one example of this failure of foresight (Dougherty, Reference Dougherty2022). The “special military operation” announced by Putin on 24 February 2022 was widely predicted by both Russian and Western experts to quickly overwhelm Ukrainian resistance (Drozdiak & Champion, Reference Drozdiak and Champion2022; Sciutto & Williams, Reference Sciutto and Williams2022). Three years later, Russia has yet to meet its military and political objectives, while suffering immense harm to its economy, international reputation and citizenry.Footnote 8
Can AI-driven decision-support systems (DSS) enhance the quality of human decision-making around the resort to force? The anticipated value of this technology includes its capacity to rapidly collect and analyze big data and uncover patterns and trends in datasets. Machine learning is also valued as a predictive tool, clarifying available and likely outcomes and calculating the risks and advantages of potential courses of action (United States Department of Defense, 2018). These are important benefits. Determining the prudence of violence on the international stage is far from simple. Decision-makers must weigh the consequences of action and inaction, the likely response of the targeted and the probabilities of escalation. Those who make these estimations routinely do so in the absence of sufficient and reliable data and sound tactical/strategic, legal and psychological counsel. If AI DSS can help policymakers and commanders escape this mire, even partly, by bridging the epistemic gaps that typically impair the quality of decision-making, should they not be welcomed?
The value of AI and machine learning also includes their potential to temper the human impulses that commonly impair effective decision-making. In their 1983 statement on Just War, the United States Bishops argued that the purpose of the “reasonable prospect of success” criterion was to “prevent irrational resort to force or hopeless resistance when the outcome of either will clearly be disproportionate or futile” (1983, p. 20, my emphasis). It is not enough to be justifiably aggrieved or righteously motivated – violence must be a rational course of action. Here again, the record of human agents makes for bleak reading. There have been a multitude of reckless wars waged across history for reasons other than instrumental (mis)calculation. Revenge (Löwenheim & Heimann, Reference Löwenheim and Heimann2008; Steele, Reference Steele, Lang, O’Driscoll and Williams2013), fear (Crawford, Reference Crawford2000), love (Hartnett, Reference Hartnett2024) and frustration (Ghalehdar, Reference Ghalehdar2021) – each of these emotions has driven imprudent and immoral uses of force at the state level. Unburdened by historical grievance or nostalgia, immune to the pressures of “groupthink” (Payne, Reference Payne2021, pp. 138–139), and free of passionate impulse, AI-driven support systems may help us better identify and reject conflict that cannot be won and therefore should not be fought.
While we should remain open to strategies and tools that strengthen resort-to-force decision-making, caution is needed regarding the suggested benefits of AI. This must begin with an accurate appraisal of the serious and not-yet surmountable limits of these technologies. One way AI might enhance “prospect of success” evaluations is via their superior quantitative reasoning; specifically, improvements to intelligence gathering, processing and retrieval. But these models will only be as good as the data they are trained on; biased datasets will perpetuate or intensify bias at the analysis phase (Cooper, Reference Cooper2024). As the evidence shows, generative AI technologies constructed on large language models are prone to “hallucinations,” misleading or erroneous translations disconnected from the source material (Sun, Sheng, Zhou & Wu, Reference Sun, Sheng, Zhou and Wu2024). AI opacity is another challenge – in some instances, the unreliability of AI DSS may be as severe as the human alternative, but more difficult to locate and correct.
We also need to ask ourselves a deeper question. What does “good” data even look like in this context? There is no universal scale of military “success.” Rather, the standard of reasonable prospect will be case specific – shaped by the social, cultural and political makeup of the actors involved and the cause and stakes of the dispute. Take Ukraine, for example. Whether this just party can succeed ultimately depends on what the term is understood to mean – the retaking of Crimea, the recovery of all Ukrainian territory seized by Russia since 2022, or survival as an independent political actor? The same is true of all other conflicts. The threshold for defeat and success is so vastly diverse, across time, between states and even at different junctures of a single conflict, that comparisons across examples may be irrelevant. The idiosyncratic nature of “success” will invariably undercut the predicted utility of AI DSS to recognize patterns in data, predict scenarios and recommend effective courses of action.
The limits of probabilistic AI in this setting become even starker when we consider the full range of factors that typically influence military success and failure in war. The Global Firepower Ranking provides an analytical display of data for 145 militaries, relating to military size and equipment, finances, geography and natural resources (Global Firepower Ranking, 2006). These material factors are important determinants of successful war-making and can, at least potentially, be better recorded and analyzed using AI technologies. What cannot be forgotten, however, are the unquantifiables of war – esprit de corps, will, luck and chance. These intangible determinants of victory and defeat necessarily frustrate efforts to assign a percentage value to the prospect of success in war. But if ignored, in favor of an exclusive, AI-assisted focus on measurable material indicators, our ability to effectively and dispassionately evaluate the prudence of Just War will only worsen. Alongside this is a deeper concern, that AI will lead us not only toward the wrong answers in matters of violence, but away from the right questions.
2.2. Broken questions
As a reminder, the aim of this article is to evaluate, through the lens of Just War, the value and drawbacks of AI in the context of resort-to-force decision-making. The fracture points of this tradition – the long-standing and enduring debates over the killing and dying of war – are relevant only so far as they inform us as to the likely ethical impact of this technology. They do this most, I argue, in matters of disposition. There are better and worse ways to morally engage war, dispositions more and less attuned to its complexity and ugliness. AI-assisted reasoning is most likely to incentivize and calcify the worst of these.
Just War is at its best when it remains a dialogue, taking war as a tragic condition as its starting point. Over the past two decades, some “revisionists” have sought to shift Just War from a tradition into a theory, substituting open-ended questions with explicit answers (Brown, Reference Brown, Hom, O’Driscoll and Mills2017, p. 86). These answers are typically derived through high-level abstraction and hypothetical, sometimes far-fetched, cases. One risk of this approach is sanitization (Lazar, Reference Lazar2017, p. 38), the unmooring of war from its inescapable and morally relevant features, such as misery, joy, terror and trauma. Disconnected from these brute facts, justice may be wrongly understood as a measurable and solvable set of problems. At its worst, this approach transforms Just War into a box-ticking technical exercise; a “decision-support system that allows us to shrug off responsibility for consequential decisions by submitting to strict criteria and calculations” (Erskine, Reference Erskine2024, p. 185).
Every element of Just War is shaped by the approach we adopt, including the principle of “reasonable prospect of success.” While the term “victory” generally implies an end to hostilities and settlement of the relevant dispute, the precise nature of this end varies radically from conflict to conflict. Deterrence restored, armies vanquished, cities captured, populations extirpated or enemies remade into allies – all have served as a marker of “victory.” The “reasonable prospect of success” principle, in most cases:
[F]unctions as little more than a rump utilitarian backstop designed to guard against feckless military adventurism. As such, it assumes rather than interrogates the concept of success and is thereby symptomatic of the general failure of just war theorists to problematize victory and its relation to just war (O’Driscoll, Reference O’Driscoll2020, p. 191).
AI-driven systems have the potential to intensify this failure by pushing us further toward an abstract, technocratic vision of military “success” empty of moral and political wisdom. War as an engineering exercise. Like most of our discussions on AI, history can serve as a useful guide when reflecting upon these dangers. The belief during the Vietnam War, for example, that the enemy and conflict would follow a prescribed formula, and thus be amenable to total control, helped undo the war effort of the United States:
Estimating the probability of success for South Vietnam and the United States proved to be the downfall of the American decision-makers … the United States underestimated the will and staying power of the North Vietnamese and Vietcong leadership and the loyalty and durability of their rank and file (O’Brien, Reference O’Brien1981, pp. 93–94).
American decision-makers over-relied on irrelevant and often inaccurate statistics, while dismissing the centrality of Vietnamese history and culture in the cause and resolution of the conflict (Adas, Reference Adas2006, p. 299). Underpinning both errors was a narrow theory of victory that elevated violent punishment as the preferred means and enemy body count as the preferred ends. There was no technological silver bullet capable of transforming defeat into victory in Vietnam. What was needed was a dispositional shift – a less grandiose political vision for the region and a prioritization of the rights and dignity of the Vietnamese people.
What about the 2011 intervention in Libya? The military outcome of this struggle between the dissolving Gaddafi regime and a multi-state NATO coalition was never in doubt. AI-enabled DSS, had they been sophisticated enough at the time, would have surely served to confirm the certainty of a swift and categorical victory. But much depends on our precise criteria for success. Perhaps NATO states ought to have expanded their imagination beyond the immediate military contest to include conditions less likely to lead to the political disintegration of the country and eventual creation of slave markets on the Libyan coast (BBC, 2017).
If “reasonable prospect of success” is to matter, morally and politically, it must be measured against the standards of a just peace. Jus post bellum is a critically important, and too often ignored, aspect of Just War, addressing the termination and post-war phase of hostilities. While jus post bellum lacks the coherent principles of the other Just War pillars, there is consensus regarding the need to extend a minimal level of dignity to the defeated. For Walzer, legitimate wartime ends must include the restoration of the peaceful status quo. (Walzer, Reference Walzer2006, p. 121). Orend expands this to include respect for the rights and traditions of the defeated and decent treatment of the innocent (Orend, Reference Orend2006, pp. 180–181).
Some may find these obligations insufficiently demanding, others too onerous. Some may reject the basic teleological assumption that the purpose of war is to secure a just peace. However one lands on this issue, what is critical is that we actually think about what we are doing; think meaningfully about the intended and foreseeable ends of our violence. What is a just peace and what should it cost are moral and political questions, not technical ones. My fear is that the technological automation of key parts of decision-making over the resort to force will delude us into thinking the opposite. Many of the most important questions of war lack definitive answers. Ethics must reflect this and remain an open-ended and deliberative practice, resistant to technocratic co-option (Schwarz & Renic, Reference Schwarz and Renic2024). As we have witnessed throughout the history of war, when questions of efficiency become a surrogate for ethical decisions and choices, our reason and our morality invariably suffer. This is the lure of AI systems. The technology promises a more optimized command over war but entails a wishing away of the messy complexities of starting and ending one.
3. AI and the probability of escalation
One potential objection to the concerns I raise above is that I am asking too much of AI. AI and machine learning systems, however sophisticated, are unlikely to ever be the primary determinant of whether or not to commence organized hostilities at scale. More likely is the use of algorithmic recommendations and predictions to inform discreet, temporally bounded violence – force short of war.
Jus ad vim is a set of moral principles governing the decision to use “limited force.” This may include acts such as a temporary no-fly zone, decapitation strikes by missile, drone, or crewed aircraft, or a Special Forces raid. In his recent book, Just and Unjust Uses of Limited Force, Daniel Brunstetter argues that the “guiding principle” of jus ad vim must be a “presumption against escalation,” given the “totalizing and unpredictable consequences of widespread conflict” (Reference Brunstetter2021, p. 153). Limited force, in other words, should not be conceptualized as a rung to be climbed on the way to a full-scale armed conflict. It is a distinct and self-contained episode of force – lesser violence in the stead of greater violence.Footnote 9
Calculating the “probability of escalation” is of critical importance when determining the just or unjust status of limited force. There are countless examples throughout history of smaller military engagements cascading into larger conflagrations of violence. For Clausewitz and others, the reciprocal escalatory dynamic between adversaries meant that there could be “no logical limit” to an act of force (Reference Clausewitz, Howard and Parat2008, p. 77). This speaks to the inertial power of violence – its tendency to evolve in unanticipated ways and intensify to undesirable levels.
Can AI-systems better calculate, and assist us in lowering, the probabilities of escalation in the context of limited force? The capacity of these systems to analyze huge volumes of data (at speeds unachievable by humans) to predict key strategic variables, including the potential response of adversaries, would be one obvious way to enhance resort-to-force decision-making. Also worth considering is the lack of emotional intelligence within these same systems. There is a large body of research showing that integral emotion inputs to decision-making can interfere with and override otherwise rational courses of action (Lerner, Li, Valdesolo & Kassam, Reference Lerner, Li, Valdesolo and Kassam2015), particularly in conditions of high uncertainty (Mosier & Fischer, Reference Mosier and Fischer2010). Panic, frustration, terror, rage and hatred may all serve to drive destructive escalation spirals. Perhaps AI can be utilized to not only reduce the uncertainty of decision-making, but also tame what Union General Tecumseh Sherman termed, “the mad passions of men” (Sherman, Reference Sherman1975, pp. 600–602). Holmes and Wheeler are among those to recognize this potential. Considering the “nuanced interplay” between AI and human decision-making in the hypothetical context of nuclear crisis management, they argue that a fusion of data-driven AI insights and human emotional intelligence and creativity may cultivate the empathy and understanding needed to lessen the risk of spiral-model driven conflicts (Holmes & Wheeler, Reference Holmes and Wheeler2024).
This potential interplay between humans and machines is worth evaluating in greater depth. Human–machine teaming will likely involve a level of digital information interfacing not yet seen in the context of armed conflict. These interfaces do not simply combine the attributes of human and machine; they transform. They expand the scope of possibility for what can, and sometimes should, be done, and “habituate users into patterns of action, even to the point of compulsion” (Dieter and Gauthier, Reference Dieter and Gauthier2019, p. 6). The risk in question here is “action bias,” the tendency to favor, non-rationally, action over inaction. This bias is a common enough occurrence, frequently manifesting in innocuous ways (i.e., the inclination of football goalies to leap rather than remain stationary for penalties). When the action in question is the authorization or escalation of violence, however, this bias becomes a more serious impediment. While not a perfect parallel, the recent use by Israel of the AI “Lavender” system for target creation provides a stark reminder of these risks:
[H]uman personnel often served only as a “rubber stamp” for the machine’s decisions, adding that, normally they would personally devote only 20 seconds to each target before authorizing the bombing – just to make sure the Lavender-marked target is male (Abraham, Reference Abraham2024).
Human–machine teaming is frequently presented as a way to elevate the virtues of both, and in relation to some aspects of war this will likely be true. Human decision-makers will draw on AI systems, at their discretion, to know more, act faster and, where destructive impulse is concerned, feel less. But we should remain alert to the risk of disempowerment. Time-compressed, opaque systems that steer rather than sharpen the gaze, and human operators rendered down into morally and politically inert actioning devices (Schwarz, Reference Schwarz2020; Sharkey, Reference Sharkey2018).Footnote 10
4. Navigating our AI future
Techno-optimism, great power arms racing, and profit motives virtually guarantee that AI will be further integrated into resort-to-force decision-making. The character and pace of this integration, however, are not yet known. I conclude this article with some recommendations as to how we might curtail the worst excesses of these technological developments:
1. Prioritize resilience in systems of resort-to-force decision–making: Underpinning much of the optimism over AI and machine learning in resort-to-force decision-making is an exaggerated faith in optimization. Through optimization, proponents of AI hope to identify and eliminate inefficiencies and streamline decision-making processes – not unimportant when the merits and risks of violence have to be calculated in time-sensitive circumstances. As we learned during the COVID-19 crisis, however, when our “just enough, just in time” global supply chains came undone, optimization is intrinsically brittle. If we over-optimize our resort-to-force decision-making, these systems will be vulnerable when conditions shift in unpredictable ways (as they predictably will). For systems to be resilient in the face of change, some degree of “slack” must be maintained. In the context of resort-to-force decision-making, this translates into the maintenance of deliberative space, where political, legal and moral considerations can be properly weighed and trade-offs balanced. This will mean inefficiency, but of the meaningful sort, allowing human agents to innovate when faced with novel challenges that confound our algorithmic tools.
2. Test for moral and political suffocation points: Testing resort-to-force AI systems against all possible scenarios that may arise after deployment is likely impossible. More attention can and must be paid, however, to identifying the points at which “meaningful human control” (MHC) can no longer be exercised within the decision chain. “Control” should be understood broadly in this context, to include the inclination to ethically reason and the capacity to ethically intervene to override AI systems where necessary. Much has already been written on the nature and limits of MHC, including within the United Nations Convention on Certain Conventional Weapons (CCW), Group of Governmental Experts (GGE) meetings held since 2013. So far, however, debate over MHC, typically understood as the preservation of human judgment and input, has centered primarily on the use of autonomous systems in battle. Explicit study is needed at the level of resort-to-force decision-making, to clarify which human-machine interfaces best preserve the agency of users and which habituate problematic patterns of action.
3. See war for what it is: The integration of AI into resort-to-force decision-making, at the level of both jus ad bellum and jus ad vim, is framed by some as a humanizing development. Flawed humans will be augmented, inefficiencies eradicated and processes streamlined, enabling more effective and ethical uses of violence. Underpinning much of this pursuit is a frustration with, and in some cases, denial of, the deep and mostly irresolvable difficulties of war. Victory is a contested concept; peace is fraught and fragile and constituted by moral and political trade-offs; conflict escalation is unpredictable and contingent. There exists within too many accounts of military AI a “resilient numbness” to these “brute realities of warfare” (Endberg-Pederson, Reference Engberg-Pedersen2023, p. 152) – a delusion that the intractable dilemmas of armed conflict can be remade into manageable problems and “solved” once the right algorithm is developed. These efforts distract us, and sometimes preclude us, from cultivating the moral and political wisdom needed to recognize and effectively balance our responsibilities over the use and resort to force.
Funding statement
No external funding.
Competing interests
No competing interests.
Neil Renic is a lecturer in military ethics at the University of New South Wales, a fellow at the Centre for Military Studies at the University of Copenhagen, and member of the International Committee for Robot Arms Control (ICRAC). He is also the Associate Director of the Military Ethics Research Lab and Innovation Network (MERLIN). Renic specializes in the changing character of war, the ethics of killing, and emerging military technologies. He is the author of “Asymmetric Killing: Risk Avoidance, Just War, and the Warrior Ethos” (Oxford University Press 2020).
Open access
