29 Aug Automatizing Patterns of Conduct: Can Artificial Intelligence Help Commanders Better Comply with the Principle of Distinction?
[Andrea Farrés is a young international lawyer specialized in IHL, international security and human rights issues.]
With the fog of war getting thicker and thicker, commanders and politicians are naturally inclined to search for tools to get guidance on how they can better comply with the international humanitarian law (IHL) targeting principles, specifically the principle of distinction.
To distinguish a civilian from a combatant, or a person who is taking direct part in the hostilities, analysing patterns of conduct has become crucial, as like the article 44(3) of the Additional Protocol I of the Geneva Conventions acknowledges, “there are situations in armed conflicts where, owing to the nature of the hostilities an armed combatant cannot so distinguish himself.”
With the use of new technologies in the rise, artificial intelligence (AI) is presented as good at pattern recognition, which could lead to the conclusion that using this tool commanders would increase their accuracy in the process of selection of targets. However, to what extent can we automatize patterns of conduct to comply with IHL? Would it be advisable to leave entirely to algorithms the decision on who is a combatant? Bearing these questions in mind, this post opts for a human-machine teaming when it comes to applying the principle of distinction.
It does so, first by acknowledging the proliferation of AI in the battlefield, explaining the cases of the SKYNET and the URSA programmes; second, by analysing how the principle of distinction is laid out in IHL, and if it provides guidance for developing the automatizing of patterns of conduct through AI; and finally, by discussing if AI is fit for this purpose, by highlighting some opportunities and challenges worth keeping in mind while developing this new technology.
Proliferation of AI in the battlefield
Despite fully autonomous weapons, popularly known as killer robots, are far away from existing, autonomy in the battlefield is a much more established reality than one may assume at first. Hundreds of a wide range of autonomous weapon systems, including “stationary turrets, missile systems, and manned or unmanned aerial, terrestrial or marine vehicles,” with different autonomy levels (for instance some of them using AI and others not) are already deployed in the battlefield. Thus, a wide range of warfare tasks are assisted by technology.
For the purposes of this post, I focus only on the systems which incorporate AI, as those are the ones which can automatize pattern recognition. Despite having access to only a handful of those types of programmes, the SKYNET or the Urban Reconnaissance through Supervised Autonomy (URSA), developed by the United States Department of Defense, serve as good examples or this technology.
The SKYNET consists of an AI machine-learning algorithm allegedly used to analyse the cellular network metadata of millions of people in Pakistan to identify couriers carrying messages between Al-Qaeda members, and thus, through the creation of patterns related to the use of mobile phones, rated the likelihood of being a terrorist. This technique of cell phone data tracking, which was first tried by the US NSA in Afghanistan, has also been used in other conflicts, such as in Yemen. However, a report based on leaked documentation argued that a flaw in how this algorithm is trained to analyse such data allegedly produced unsound results. When analysing how the SKYNET works, its supporters argue that through programmes like this human control is not jeopardized because this is one of the multiple tools used which help the commander to reach a better assessment of who is a combatant. Besides, it enables the collection of massive amounts of information to identify leads for a targeted attack, something which through human capabilities only would not be possible. However, its critics emphasize the lack of relevant data (“ground truths”) which can be inserted to the machine and which could eventually lead to unsound results. Moreover, they stress how the process of data interpretation is not a neutral task, and how biases play an important role in developing new technology. This is so as, for instance, patterns of conduct which would increase the rate of a person to be considered as terrorist according to SKYNET would be conducts such as a low use of your phone, or switching it off, which would be automatically considered as attempts to evade mass surveillance.
The URSA programme aims to “enable improved techniques for rapidly discriminating hostile intent and filtering out threats in complex urban environments”. In other words, the DoD is “developing a program of high-tech cameras mounted on drones and other robots that monitor cities, which enable identification and discrimination between civilians and terrorists through machine learning computers.” In this second example too, one can observe the importance of identifying patterns of conduct, especially in urban contexts, to separate civilians from combatants. However, while in the SKYNET programme AI is aimed at a very specific task (rating the likelihood of someone being a terrorist analysing a pattern of conduct through the use of your phone), the URSA programme has a much broader goal, which comes with bigger challenges and opportunities. In this case, the defenders of this technology argue as well that this is one additional tool that assists commanders, not jeopardizing the requirement of meaningful human control. Moreover, URSA enhances human capabilities, as through a complicated system of sensors and AI, separates innocent from hostile behaviour, processing huge amounts of data otherwise inaccessible. However, thorny questions also arise: which behaviours are deemed as hostile or innocent? Is cultural sensitivity being considered? How to ensure that the programme is fed with accurate information? Who is monitoring the process of data interpretation? Unfortunately, as the development of the programme is confidential, these questions remain unanswered for the general public.
Can the Principle of Distinction Laid out in IHL Guide the Elaboration of Patterns Through AI?
In relation to International Armed Conflicts (IAC), article 50 of the AP I to the Geneva Conventions describes the category of civilian population, who cannot be the object of any attack, in a negative sense. To determine who belongs to the armed forces, article 44(3) of AP I abolishes the precondition of “having a fixed distinctive sign recognizable at a distance” in “situations where, owing to the nature of the hostilities, an armed combatant cannot distinguish himself from the civilian population”, being limited to the obligation of carrying his arms openly. This paragraph contains an exception to the fundamental requirement of combatants to distinguish themselves, a provision highly contested among States at the Diplomatic Conference. Even though in non-International Armed Conflicts (NIAC), common article 3 to the Geneva Conventions does not define the term “civilians”, the ICRC 2009 Guidance clarified that its definition was the same applicable for IACs. Therefore, as the category of civilians is constructed in a negative sense, there is no definition which can be inserted in the AI programmes articulating what constitutes a civilian pattern of behaviour.
In the same vein, IHL provides little guidance on which behaviour amounts to civilians directly participating in the hostilities (DPH). If a civilian is found DPH, the conduct he or she undertakes “suspends his or her protection against the dangers arising from military operations,” although no definition of what a direct participation is can be found in IHL. To assess if a civilian is DPH, the ICRC clarified that the threshold of harm, direct causation and the belligerent nexus needs to be proved. To analyse these three requirements, contextual information, “the big picture”, needs to be taken into consideration. For instance, “the tactical and strategic implications of a potential harm; the status of other potentially threatened individuals; the direct causal implications of someone’s actions; or the sociocultural and psychological situation in which that individual’s intentions and actions qualify as military actions.” This means that, as an example, feeding the algorithm with the ground truth that “driving a truck full of ammunitions” is a conduct which amounts to DPH would be incorrect, as this would be true only if the truck is going to a shooting position, and not from a factory to a port far from a conflict zone, requiring taking into consideration contextual information, a capability out of reach for AI programmes.
Using AI to Increase Compliance with IHL: Challenges and Opportunities for Human-Machine Teaming
The SKYNET and URSA programmes are a few of the examples relatively open to the public which use AI to create patterns which aim at a better compliance with the principle of distinction. Bearing these examples in mind, and how the principle of distinction is laid out by the Geneva Conventions, this post concludes by highlighting some of the challenges and opportunities automatizing patterns of conduct could create.
When considering how to automatize a pattern of conduct through AI, the first challenge one encounters is that the Geneva Conventions do not provide any hint on how a civilian behaves or what he or she looks like: they are described by opposition and in a negative sense, leaving us with no information related to the identification of civilians based on their behaviour. In relation to this, cultural sensitivity while coming up with patterns of conduct is a very much needed skill. This is so as, for instance, in various parts of the world civilians carry weapons for self-defence purposes, a pattern of conduct a Western technologist could easily qualify as linked to the hostilities if he or she has no relevant training related to the areas in which the developing technology plans to be applied. In this regard, it is important for the developers of these new technologies to bear in mind the need to be culturally inclusive and ensure that the reality on the ground is understood, so the ground truths inserted in the algorithm enable the protection of civilians.
Creating patterns of behaviour on what constitutes a DPH act or not also entails the risk of leaving out of the assessment the “big picture”, a constraint deemed necessary by the ICRC when evaluating the threshold of harm, direct causation and the belligerent nexus. Also, IHL states that solely the presence of military or civilians DPH among the civilian population does not deprive the population of the protection from an attack, favouring the need for military commanders to issue context-based decisions. Furthermore, identifying hors de combat and combatants or civilians DPH surrendering also require a contextual analysis, and the ability to interpret human intentions. As previously stated, AI inherently lacks both capabilities, which according to IHL are necessary. Therefore, when dealing with human-machine teaming to analyse the patterns of conduct created by AI, the role of the commanders is to study this data while ensuring a meaningful human control by bringing in the contextual information and the interpretation of human intentions.
On another level, there are technological challenges related to AI which hinder an accurate automatization of patterns of conduct, like the need for extensive databases to feed the algorithm with, or the automation bias, which refers to the natural inclination of humans to believe what the computer suggests, jeopardizing meaningful human control over targeting decisions. As the critics of the SKYNET pointed out, machine-learning needs massive amounts of data to come up with accurate patterns. However, there seems to be not enough evidence on “how a terrorist behaves” in order to produce sound results. Nevertheless, this challenge could be turned into an interesting opportunity to gather experts from several fields to discuss what are conducts which amount to DPH, so we could move from having a handful examples to draw on, to have sound generalizations which could be welcomed both at the legal level and would facilitate the development of patterns through AI.
In relation to the opportunities the use of AI offers, the potential reduction of the confirmation bias is an interesting option. This bias, which is present not only when AI is used, as it is a typical characteristic of human thinking. It refers to when a person has a certain belief (for example, that X behaviour is typical of a person who is a combatant), so therefore he or she is going to focus on all the clues that confirm his or her belief, omitting the opposing information that would challenge this perception. Also, this is more likely to happen when little evidence is available, which can often be the case, as it has been demonstrated by some studies related to wrongful drone strikes. In this regard, the development of AI could reduce the confirmation bias if used as an added tool for the commander to agree on a targeting decision, as it would enable him or her to work with more information, which by herself of himself would be unable to process, and because the recommendation made by the algorithm would be the collective result of the work of lawyers, roboticists and technologists, suggesting a result not linked to the certain beliefs of the commander.
In conclusion, one should bear in mind that although AI can be better than humans at generating information, human capabilities to analyse it are still much accurate than the ones of a “robot”. Therefore, while developing human-machine teams, the existence of biases should be taken into account, as well as the limitations current AI capabilities have, such as the lack of contextual awareness or little information available from which to draw reliable patterns.