The horrors that warfare can impose on soldiers, fighters and on hapless civilians unwillingly caught up in the fight was played out in the 1940s throughout Europe, North Africa, in the Atlantic and Pacific Oceans and Far East, in the 1960s and 1970s in Indo-China, in the 1990s in the former Yugoslavia, in more recent times in Libya and Syria and in all of these years and many more besides in countless other parts of the world. As carefully worded legal rules strive more prescriptively to protect those who try to keep out of the fight, it is those very individuals who form an ever-increasing proportion of the casualties. War in which men use weapons against one another is, truly, a miserable affair.

Against this depressing reality, in this article we consider whether increasing levels of automation in attack and futuristic notions of autonomous attack decision-making represent an opportunity to be grasped with both hands, or yet another unwanted scientific advance that promises even less discrimination in the ‘killing game’.

Autonomy and automation are of course discrete notions, but we must be clear as to the meaning to be ascribed to each. Section 4.2 of this article will therefore assess what a clear taxonomy might look like. International law prohibits the use of some weapons, means and methods of warfare in armed conflict and imposes restrictions on the use of other weapons. In Sect. 4.3 we will consider the weapons law rules of most evident relevance to autonomous attack technologies. A distinct branch of the law of armed conflict, the law of targeting, regulates how weapons may be used. In Sect. 4.4, we will discuss what targeting law rules are likely to have the greatest relevance to autonomous attack technologies. The notion of autonomous attack decision-making has provoked controversy, and Human Right Watch has proposed a comprehensive ban on such technologies. In Sect. 4.5 we will ponder whether a comprehensive ban is merited at this stage. In Sect. 4.6 we will conclude by proposing a way ahead.

The 2002 attack in which the US targeted Qaed Senyan al-Harthi in Yemen by means of a Predator remotely piloted aircraft equipped with a Hellfire missile proved the concept of aerial attack in the modern era using remotely piloted aircraft.¹ Applied to this use of the air environment, automation involves the mechanisation of the platform’s decisions to a degree falling short of autonomy. It would, however, be wrong to think of these matters exclusively in terms of the air environment, as automation and autonomy are equally applicable in maritime warfare,² on land, in outer space and indeed in cyberspace. Nevertheless, elements of the discussion in the present article will be by reference to air platforms. But what do automation and autonomy mean? We start by considering what doctrine has to say.

An automated system has been described as one that, responding to inputs from one or more sensors, is programmed logically to follow a pre-defined set of rules to provide an outcome.³ If you know the rules under which it operates, you can predict that outcome. An automated system functions in a self-contained manner once deployed, and will independently verify or detect a particular type of target object and then fire or detonate the munitions.⁴ Automated weapons are nothing new; consider for example certain kinds of mine or booby-trap.⁵

Autonomous systems, by contrast, take the matter several stages further. They employ an understanding of higher-level intent and direction and an awareness of their environment to take appropriate action and thereby bring about a desired state. Critical to autonomy is the system’s ability to decide a course of action from alternatives without depending on human oversight and control. Its overall activity is predictable but individual actions may not be.⁶ The system operates independently; the software identifies and engages targets without being programmed to target a specific object. Importantly, the International Committee of the Red Cross (ICRC) has suggested that development of a truly autonomous weapon system that can implement international humanitarian law represents a monumental programming challenge that may well prove impossible.⁷

No doubt the increasing role of automation in everyday life and consequent familiarity with it will influence its perceived acceptability in warfare, and logic suggests that peacetime technologies that are clearly adaptable for use in military environments will increasingly be seen on the battlefield.⁸ This will be attributable to their appeal, which in turn will be derived, for example, from their convenience of use, from their tendency to limit the dangers faced by friendly force personnel and from their ability to enable timely responses to rapidly appearing, potentially overwhelming threats. The future battlespace will involve threats that require rapid and flexible responses at speeds that presuppose automated decision-making; other advantages claimed for these new technologies include saving lives; fearlessness; remembering orders; absence of emotional responses; suitability for dull, dangerous and dirty tasks; needing no or less rest; having shareable intelligence and computation speed.⁹ It follows from this that the weapons discussed in this article are likely to be seen as having such utility that their further development and eventual procurement will likely be regarded as critical to operational success.

Autonomous weapons can loiter, seek, identify and engage targets and can report the point of weapon impact. The Wide Area Search Autonomous Attack Miniature Munition, for example, is a small cruise missile with a loiter capability that can seek a specific target and that, on acquisition, attacks or requests permission to do so. The autonomous element of the weapon is, it seems, posing significant engineering issues, for example because it is likely beyond current technology for the machine to make the complicated assessments required to determine whether or not a particular attack would be lawful if there is an expectation of collateral damage.¹⁰

Burrowing down a little deeper, “so long as it can be shown that the system logically follows a set of rules or instructions and is not capable of human levels of situational understanding, then they should only be considered automated.”¹¹ If that is doctrine’s interpretation, Peter Asaro describes as autonomous “any system that is capable of targeting and initiating the use of potentially lethal force without direct human supervision and direct human involvement in lethal decision-making,” a formulation which could include systems current doctrine sees as ‘automated’. So working out a widely accepted taxonomy is a clear priority for future work.¹²

For example, under national doctrine, an unmanned aircraft would be part of an automated system if, first, it is pre-programmed either to proceed to a set location and there to fire a weapon or if, second, having reached a pre-set location, it searches a defined area of territory for specified objects which, when detected, it recognises using on-board image recognition technology, and attacks. If, third, a similarly equipped unmanned platform were additionally programmed to make evaluations or decisions or to undertake certain procedures before it then decides whether and how to undertake the attack, this would seem to render it an autonomous as opposed to an automated system. So the doctrinal distinction seems to lie in genuinely mechanical decision-making processes that go beyond simple recognition resulting in automatic attack.¹³

An apparently different understanding is reflected in a recent United States Department of Defense Directive¹⁴ which describes a weapon system that “can select and engage targets without further intervention by a human operator” as autonomous. It would seem clear that this would characterise the second and third examples discussed in the previous paragraph as autonomous systems.

The US definition would seem to focus the discussion on all technologies that cause the machine to select the target for an attack without human involvement at that moment. It seems to matter little to this definition that recognition technology is used to decide that an observed object or person is a target. The point deserves repeating that if we are to have a sensible debate on these matters, generally accepted terminology is essential. For the purposes of the following discussion, autonomy and automation will be understood in accordance with the JDN 2/11 definitions noted earlier in this Section. Legal issues are likely to arise, however, in relation to both autonomous and certain automated attack technologies, so in the following discussion the reader will frequently see reference to both.

One of the core, customary rules of the law of armed conflict requires that a distinction be constantly maintained between civilians and combatants and between civilian objects and military objectives.³⁴ Whether autonomous attack technologies can be employed consistently with this principle will depend on the technical performance of the recognition technology. Where there is doubt as to certain matters, an attack must not proceed.³⁵ In the case of attacks against objects, the decisive question is whether a weapon system can differentiate sufficiently between, for example, the military objects³⁶ it is designed to recognise and civilian objects that are protected by the law. An automated or autonomous system may, for example, be programmed to examine an object it observes by reference to the particular characteristics of, say, a tank, artillery piece or armoured personnel carrier. If sufficient points of similarity are achieved, the controlling software may determine that the observed object is a military object that it is lawful for the weapon system to attack. Testing will, however, be important to evaluate the reliability with which the weapon system conducts this recognition process and, thus, its ability to comply acceptably with the distinction principle. Such testing may prove challenging and it seems likely that computer modelling will also be required.³⁷

Even greater technical challenges confront the development of technology for the automated or autonomous attack of individuals. The principle of distinction, when applied to attacks that target persons, requires that combatants and civilians directly participating in hostilities be differentiated from civilians taking no part in the hostilities.³⁸ Some research may focus on the mechanical observation of characteristics peculiar to combatants, such as their metallic footprint or, perhaps, aspects of their behaviour and movement.³⁹ In relation to metallic footprint, however, nanotechnological developments in military equipment such as the manufacture of rifles using plastics and other substances may challenge the appropriateness of such an approach.