The idea of “compassionate AI” in warfare appears paradoxical at first glance, yet it becomes analytically meaningful when defined with precision. For the purpose of this article, “compassionate AI” refers not to emotional sensitivity or moral intuition, but to the technical capacity of an autonomous weapon to apply the central principles of international humanitarian law (IHL), especially distinction and proportionality. These principles, together with precaution, humanity, military necessity, and the prohibition of unnecessary suffering, form the normative architecture that governs the conduct of hostilities (Henckaerts & Doswald‑Beck, 2005).
Recasting “compassionate AI” in these terms exposes the profound conceptual, legal, and practical challenges of delegating lethal decision making to machines (Sharkey, 2012). It also reveals a deeper truth: the future of autonomous weapons systems (AWS) depends not on compassionate machines, but on compassionate human institutions capable of guiding, supervising, and constraining these systems in their development and their application phase (Asaro, 2012).
To imagine an autonomous weapon capable of applying distinction is to imagine a machine capable of recognising, in real time, the difference between a civilian and a combatant, between a human shield and a legitimate target, between a civilian object and a military one. Distinction is not a matter of pattern recognition alone; it requires contextual interpretation, often cultural understanding, and certainly the ability to infer intent from behaviour (ICRC Commentary to Additional Protocol I, 1987). Any war crimes trial records would indicate that even the human creators of IHL more often than not have difficulty making these distinctions and decisions in real time in the heat of the battle and in an act of self‑preservation or sheer desire to be victorious (ICTY, Prosecutor v. Galić, 2003).
Proportionality is even more demanding. It prohibits attacks in which the expected incidental civilian harm would be excessive in relation to the concrete and direct military advantage anticipated (Additional Protocol I, Art. 51(5)(b)). This is not a mathematical calculation but a moral and legal balancing act, requiring judgement about the value of human life, the significance of military objectives, and the acceptability of risk (Boothby, 2012). Precaution adds a further layer of complexity: all feasible measures must be taken to verify targets, minimise civilian harm, and cancel or suspend attacks if circumstances change (Additional Protocol I, Art. 57). These principles are constrained by humanity, which prohibits unnecessary suffering and insists that even in war, human dignity must be preserved (Martens Clause, 1899).
Military necessity, often misunderstood as a licence for violence, is in fact a limiting principle: it permits only those measures that are necessary to weaken the enemy’s military forces and are not otherwise prohibited by IHL. Together, these principles form a web of obligations that require interpretation, deliberation, and moral reasoning – which requires compassion (Satyarthi, 2026), defined as “mindful, action‑oriented problem solving,” or in this case mindful decision to apply the above principles. In the case of AWS, this would imply intentional international regulation (Crootof, 2015).
To encode such obligations into an autonomous weapon, a programmer would need to be simultaneously a computer scientist, an IHL specialist, and a military operational expert. They would need to understand not only how to build perception algorithms, but how to interpret the Martens Clause, how to assess proportionality in fluid environments, how to recognise when military necessity has been stretched beyond its limits, and how to ensure that no weapon causes unnecessary suffering. Such a combination of expertise is extraordinarily rare. Even if such an individual existed, the translation of open‑textured legal norms into code would remain inherently limited (Boulanin & Verbruggen, SIPRI, 2017). Law is interpretive; code is literal. Law tolerates ambiguity; algorithms do not.
Because no algorithm can fully replicate legal judgement, AWS cannot be ethically or legally deployed without continuous human supervision (ICRC, 2017; GGE LAWS Reports 2017–2024). This supervision must be exercised by individuals who understand both the legal framework and the operational context. They must be able to monitor target selection, override or abort engagements, assess proportionality dynamically, and ensure that precautionary obligations are fulfilled. They must recognise when military necessity has been invoked too broadly, when humanity requires restraint, and when the risk of unnecessary suffering is too great.
In practice, this means AWS must remain a semi-autonomous tool purely for analysing and advising, with humans retaining meaningful control over critical functions (Human Rights Watch, 2012). The idea of a fully autonomous, legally compliant, ‘compassionate’ weapon is therefore a contradiction in terms. The machine may process data, but only a human can interpret meaning.
The introduction of AWS also creates a profound accountability dilemma (Mako, 2026). When an autonomous system causes unlawful harm, who is responsible? The programmer who wrote the code may have influenced the system’s behaviour, but they do not control the battlefield, cannot foresee all operational contexts, and do not possess command authority. The military commander is responsible for target selection and ensuring compliance with IHL, yet if a commander issues a broad order, such as “neutralise enemy combatants”, and the AWS misidentifies civilians, the chain of responsibility becomes blurred (Inoyatov, 2025). Did the commander fail to provide adequate constraints? Did the system exceed its intended parameters? Did the programmer fail to anticipate a scenario? Under IHL, the state bears responsibility for the conduct of its armed forces and the weapons it deploys (Articles on State Responsibility, 2001), but AWS introduce layers of technical opacity that complicate attribution. When decisions emerge from complex interactions between sensors, models, and algorithms, the causal chain becomes difficult to reconstruct. The core issue is not individual blame but systemic diffusion of responsibility (Crootof, 2015). AWS challenge the foundational assumption of IHL – that human beings make decisions about the use of force, and the genetic principle of IHL – “Do no harm”. Applying this principle and fighting enemy combatants is supposed to be feasible for humans according to the creators of the universally ratified 1949 Geneva Conventions, but is algorithmically unattainable.
As Mako (2026) notes, landmines are among the earliest forms of autonomous weapons: once deployed, they operate without human oversight, and they continue to kill and maim long after hostilities have ended. The 1997 Ottawa Convention prohibits anti‑personnel mines precisely because they are inherently indiscriminate. Yet recent conflicts show how easily such rules can be ignored in practice, and how readily states may violate the very international conventions to which they have formally acceded – not only in armed hostilities but also in efforts to deter or halt irregular migration.
Accountability challenges are compounded by the structural vulnerabilities of AI‑enabled targeting: such systems may misidentify civilians as combatants due to biased or incomplete training data, adversarial environments, sensor limitations, or contextual ambiguity (Kroll et al., 2017). These risks are not hypothetical. They are widely discussed in relation to modern targeting systems, where misidentification can have tragic consequences (Schmitt & Thurnher, 2013). The question is not whether a specific company or software is responsible for specific incidents, but how the underlying architecture of AI‑driven targeting can fail to distinguish a civilian from a combatant.
When such failures occur, they are not simply technical errors – they are legal, ethical, and institutional failures (ICRC, 2026). They represent a breakdown in the application of distinction, a miscalculation of proportionality, a failure of precaution, and ultimately a violation of humanity.
Beyond technical limitations, AWS operate within human institutions that shape their use. Public testimony from military officials involved in past air campaigns has described operational pressures such as “maintaining the tempo,” ending in the expansion of target categories once clearly military targets are exhausted (House of Commons Defence Committee, 2000). These dynamics illustrate how institutional incentives can lead to civilian harm even without malicious intent. AWS deployed within such systems may amplify, rather than mitigate, these pressures. A machine does not question orders; it optimises them. If the institutional logic rewards the number of targets struck, an autonomous system may simply accelerate the pace at which ambiguous or marginal targets are engaged. Again, the problem is not the machine’s lack of compassion, but the institution’s lack of restraint.
Intentional international regulation alone will not suffice. If AWS require compassionate human supervision, then the institutions responsible for selecting, training, and overseeing those humans must themselves be grounded in compassion. The current competency frameworks of public and private sectors are lacking in this respect, and require revision of competency frameworks that integrates compassion into recruitment criteria, embeds ethical reasoning into training, and ensures that leadership pathways reward moral and compassionate judgement rather than bureaucratic efficiency. Compassion would become a structural principle, not merely an individual virtue.
In conclusion, “compassionate AI” in the sense of an autonomous weapon capable of applying distinction, proportionality, precaution, humanity, and military necessity is both technically unachievable and conceptually flawed. These principles require human judgement, contextual awareness, and moral reasoning. AWS must therefore remain semi‑autonomous, guided by compassionate human experts who understand both IHL and military operations. Ensuring accountability requires reformed institutions, reimagined competency frameworks, and a renewed commitment to human dignity. The future of AWS governance depends not on compassionate machines, but on compassionate human institutions capable of bearing the moral weight that machines cannot carry.
And because autonomous weapons threaten global strategic stability, challenge the core tenets of international law, and pose urgent ethical risks, their governance falls primarily into the domain of foreign policy and multilateral diplomacy and international standardization. They are far from constituting a purely domestic policy concern or a prerogative of private defence and security companies.
The International Committee of the Red Cross (ICRC), as the mandated custodian and interpreter of international humanitarian law, plays a central role in shaping global debates on autonomous weapons systems. In its engagements with states, militaries, and technologists, the ICRC has consistently warned that no autonomous system can replace the human judgement required by IHL, and that meaningful human control must remain at the heart of any lawful use of force.
As for compassion itself, it is to be “teached”, not preached.