29 Nov Three Individual Criminal Responsibility Gaps with Autonomous Weapon Systems
[Dr. Marta Bo is an Associate Senior Researcher on emerging military technologies at Stockholm International Peace Research Institute (SIPRI), Researcher in Public International Law at the University of Amsterdam- Asser Institute and Research Fellow at the Geneva Graduate Institute.]
To hold an individual criminally responsible for committing an unlawful attack, it must be established that they launched the attack with some form of intent or knowledge. However, what if an attack targeting civilian objects and military objectives indiscriminately involved an autonomous weapon system (AWS)? Let’s say the AWS was pre-programmed to target military targets (say a tank) but ends up targeting a bus full of civilians? Could the AWS operator be held criminally responsible for violating the prohibition of indiscriminate attacks on the basis that they failed to intervene and suspend the AWS-attack? Assuming that the AWS operator lacked intent, would it make a difference if the AWS operator failed to suspend the attack because they did not perceive the risk of targeting a civilian object or they foresaw such a risk and decided to proceed nonetheless?
These scenarios are far from unlikely in the future.
This post argues that AWS shed light and potentially exacerbate three critical gaps in the framework governing individual criminal responsibility for war crimes. These gaps or difficulties relate to the criminalization of failures to intervene in the operation of an AWS (i.e. omissions), the war crime of indiscriminate attacks, and the criminalization of risk-taking behaviours.
Failures to Intervene in the Operation of an AWS…
In debates on accountability and AWS it is important to look at whether omissions, notably failures to intervene in the operation of an AWS, could trigger the responsibility of an AWS operator for committing a war crime. Aside from command responsibility, which falls outside the scope of this post, whether omissions trigger criminal responsibility (for direct commission) is debated.
‘Human-machine interaction’ is at the core of the conversation around AWS, spurring debates over whether IHL requires human supervision and intervention during an attack (see recently McFarland). At least in dynamic and unpredictable targeting and conflict environments, there is a certain degree of consensus among some experts and states that human supervision and the possibility to intervene during an attack is a requirement for complying with IHL (SIPRI, p. 9). Human supervision is considered necessary in some situations to ensure that the user of the AWS can verify targets and suspend the attack if it is expected to be unlawful. Although it is considered necessary, there are practical difficulties of how this would work in practice given the speed and scale at which AWS may operate.
At the operational stage, AWS operators may fail to properly supervise the system and intervene when necessary. For example, during the course of an attack an AWS operator may fail to verify targets and/or to stop an attack expected to be unlawful. These failures may lead to attacks in violation of the IHL principles of distinction or proportionality. This highlights the need to evaluate failures to intervene in the operation of an AWS and its consequences. Will such failures attract the individual criminal responsibility of the AWS operator for committing an unlawful attack?
The starting point is that war crimes provisions related to unlawful attacks are formulated in terms of positive actions, reflecting the commission of an act rather than an omission to act. For instance, Additional Protocol I (Art. 85 (3)(a)-(e)) refers to ‘making’ protected individuals and objects ‘the object of attack’, ‘launching an indiscriminate attack’, ‘causing death or serious injury to body or health’, whilst the Rome Statute references ‘directing’ or ‘launching’ unlawful attacks in Art. 8(2)(b)(i)- (iv). Another problem is that failures to verify objectives and suspend attacks are violations of the IHL duty to take precautions in attacks. However, violations of precautionary measures are neither criminalised by AP I or the Rome Statute.
Is responsibility thus evaded for these modes of commission of unlawful attacks when it comes to AWS? The answer hinges on the possibility of prosecuting omissions as unlawful attacks.
Most national criminal laws accept ‘commission by omission’, in other words, that a crime can be perpetrated by omission as well as positive action. Equivalence clauses in criminal codes place an omission on an equal footing with an act ‘when the offender is under a duty to act and prevent the occurrence of an event that forms part of an offence normally committed by positive acts’ (Bohlander, p. 41). At the international level, Article 86 AP I obliges High Contracting Parties and the Parties to the conflict to repress, including by criminalizing, grave breaches of AP I resulting from a failure to act when under a duty to do so. This norm allows for the equivalence of acts and omission in the context of grave breaches of API, including war crimes of unlawful attacks. In contrast, there is greater uncertainty under the Rome Statute. Not only does the general part of the Rome Statute lack an ‘equivalence clause’, but the preparatory works also indicate that a provision on omission liability was deleted. Scholars remain divided on the implications of this for whether it criminalises crimes by omission (see Roth 60; Ambos 194; Saland, 213; Duttwiler, 57-59).
The primary condition for holding an individual responsible for committing a crime by omission is that they ‘breach a duty which required them to actively safeguard a protected interest in a specific situation’. Are AWS operators under individual duties to act under IHL? Despite debates on the direct applicability of IHL to individuals, there is a certain degree of consensus that precautionary obligations impose duties to act on individuals (plus, they are enshrined in the military manuals of the majority of states). In particular, the obligation to verify objectives is specifically addressed to ‘those who plan or decide upon an attack’, whilst the obligation to suspend has also been interpreted to create individual duties. Importantly, the latter is considered to apply to the executers of an attack, such as AWS operator, when an attack is expected to be in violation of the principle of distinction, including when attacks are expected to indiscriminately target military objectives and civilians or civilian objects (Boothby, 120). While one could argue that monitoring an attack with an AWS differs from executing an attack, this norm is arguably important for ascribing responsibility for commission by omission to AWS operators. Moreover, duties to take precautions in attacks are aimed at preventing or minimising collateral damage to civilians and civilian objects, which is the same aim underscoring the criminalization of unlawful attacks as war crimes.
Finally, ‘commission by omission’ is ostensibly limited to crimes of result. While war crimes of unlawful attacks under the Rome Statute proscribe certain conduct (such as directing an attack against civilians or launching a disproportionate attack), the parallel crimes under AP I require a particular result, i.e. ‘causing death or serious injury to body or health’. Crimes of result criminalise a certain result that must occur in addition to the underlying conduct. What is relevant is the conduct, be it an act or inaction, triggering the causal process leading to civilian damage (death or injury).
In sum, whether there is a responsibility gap for expected instances of war crimes committed with AWS largely depends on: whether a legal system allows for the equivalence between acts and omissions in the commission of war crimes, whether in its absence a state could employ Art. 86 AP I on the basis of its constitutional arrangements, and how the war crime of unlawful attacks is defined. Finally, the ascription of criminal responsibility for omission crucially depends on the individual’s capacity to perform the required act and to discharge the legal duty to act (according to most national legal systems and in the ICTY case law). As I have observed elsewhere, this raises a complex set of issues in the case of AWS.
…Resulting in Indiscriminate Attacks…
Failures by AWS operators to intervene and halt an attack with an AWS could result in indiscriminate attacks. Indiscriminate attacks are defined at Art 51(4) AP I as attacks that ‘strike military objectives and civilians or civilian objects without distinction’ because: they are not directed against a specific military objective; or because they employ inherently indiscriminate weapons, which are those that cannot be directed at a specific military objective or whose effects cannot be limited as required by IHL. As such, this prohibition specifies the principle of distinction. At the same time, AP I considers disproportionate attacks an example of indiscriminate attacks.
It is recognised that AWS could, in some instances, increase the risk of errors in target identification (SIPRI p. 27; see however Trabucco and Heller) resulting in unlawful attacks against civilians. It is debated whether accuracy and reliability in target identification are suitable proxy for distinction and appropriate criteria to assess whether an AWS could be qualified as inherently indiscriminate by nature or whether the use of an AWS would amount to an indiscriminate attack (SIPRI pp. 13 and 14).
However, while errors in discrimination do not automatically equate to a violation of the rule of distinction, the potential for such errors means that it is this rule – rather than the prohibition on direct attacks against civilians – most likely to be violated.
The problem is that war crimes provisions only partially mirror IHL prohibitions. Both AP I and the Rome Statute criminalise direct attacks against civilians and disproportionate attacks. However, neither instrument defines a war crime of indiscriminate attacks as a violation of the rule of distinction. Does this mean that the most likely violations of IHL carried out with AWS will not be covered by war crimes provisions? Is there a resulting criminal responsibility gap for one of the most likely IHL violations with AWS?
To address this question, one should look at how international criminal tribunals have treated indiscriminate attacks. Throughout the years, the ICTY has considered indiscriminate attacks to be covered by the war crime of directing attacks against civilians (Blaskić TC, §180; Galić TC, §57). The ICC more recently used the war crime of attacking civilians as a basis for prosecuting indiscriminate attacks (Ntaganda, TC, § 921).
Subsuming indiscriminate attacks under the war crime of directing attacks against civilians has been a way to overcome the gap in international criminal law with respect to the lack of a specific war crime provision. However, this is not without challenges as these two war crimes are inherently different: indiscriminate attacks refer to a ‘lack of targeting’, while directing attacks against civilians reflects the intention to target civilians (or other protected individuals or objects). The mens rea of the perpetrator of an indiscriminate attack, at least under the ICC framework, is arguably different from the one who intentionally targets civilians insofar as it consists, in most instances, of risk-taking conduct.
Lack of clarity as to the criminalisation of indiscriminate attacks and the elements of this crime could be an obstacle to ascribing criminal responsibility in likely violations of IHL with AWS and give rise to a responsibility gap.
…Stemming from Risk-Taking Behaviour…
Provided that failures to suspend attacks with AWS resulting in indiscriminate attacks are criminalised, one should then turn to the mens rea element. If an operator of an AWS fails to suspend an attack expected to be unlawful with the intention to target civilians, then this scenario is undisputedly covered by the mental elements of ‘wilfulness’ under AP I and ‘intentionality’ under the Rome Statute. Moreover, situations of indirect intent (dolus indirectus), where the user fails to suspend an attack with an AWS and is ‘practically’ or ‘virtually’ certain that the attack will be directed against civilians or result in civilian death or injuries, are covered by these mental elements. However, as already pointed out by many (Ohlin, Bo, SIPRI), due to problems of understandability of AWS compounded by the speed of military targeting, there will most probably be some forms of reckless or negligent failures to act, intervene or stop AWS. For example, an AWS operator may envisage the risk of targeting civilians who are intermingled with military objectives in a certain area, but dismiss that risk and proceed with the attack. If such a reckless failure to halt an attack with AWS result in an indiscriminate attack against civilians, does this trigger criminal responsibility?
AWS raise the oft-debated issue of criminalisation of risk-taking behaviour in targeting. While the mental element of ‘wilfulness’ under AP I has been interpreted as including recklessness, this has been excluded by the Rome Statute and there are diverging opinions as to the criminalization of risk-taking behaviour in the form of dolus eventualis, due the mens rea requirement of ‘intentionality’ under the Rome Statute as currently interpreted by the ICC (Katanga, TC 806) (see here, here).
Open interpretative debates around the prosecution of targeting decisions and attacks based on risk-taking behavior as war crimes could accentuate difficulties to ascribe responsibility for unlawful attacks with AWS.
AWS shed light on three existing gaps in the framework of individual criminal responsibility for war crimes: whether omissions and violations of the duty to take precautions in attacks could amount to a mode of commission of a war crime; the criminalization of the war crime of indiscriminate attacks; and the criminalization of risk-taking behaviours. These gaps are potentially accentuated because: those executing attacks are and will increasingly hold supervisory roles over AWS; AWS hold the potential to increase the risk of errors in the identification of targets; and, linked to that, lack of understandability of AWS compounded by the speed of military targeting may, in some instances, make AWS operators more prone to risk-taking behaviour.
Addressing these pre-existing gaps and legal uncertainties in the current system of individual criminal responsibility for war crimes is crucial to ensure accountability for expected violations of IHL involving AWS. At this particular juncture of the regulatory debate where the CCWUN adopted the guiding principle that ‘[a]ccountability for developing, deploying and using any emerging weapons system in the framework of the CCW must be ensured in accordance with applicable international law’, it is high time that this issue is given greater attention.
The framework of individual criminal responsibility for war crimes is also recognised by many as a crucial means of ensuring compliance for IHL. While it operates ex post, it provides useful insights for what should be prevented ex ante and how. As such, addressing the pre-existing gaps in this framework presents an opportunity for states and experts to elaborate on the standards of intent, knowledge and behaviour that IHL compliance demands in the use of AWS. This could help reduce the risk of at least some serious violations of IHL, such as indiscriminate attacks stemming from reckless failures to intervene and suspend an AWS.