Recent work on interpretability in machine learning and AI has focused on the building of simplified models that approximate the true criteria used to make decisions. These models are a useful pedagogical device for teaching trained professionals how to predict what decisions will be made by the complex system, and most importantly how the system might break. However, when considering any such model it’s important to remember Box’s maxim that "All models are wrong but some are useful." We focus on (...) the distinction between these models and explanations in philosophy and sociology. These models can be understood as a "do it yourself kit" for explanations, allowing a practitioner to directly answer "what if questions" or generate contrastive explanations without external assistance. Although a valuable ability, giving these models as explanations appears more difficult than necessary, and other forms of explanation may not have the same trade-offs. We contrast the different schools of thought on what makes an explanation, and suggest that machine learning might benefit from viewing the problem more broadly. (shrink)

Previous research in Explainable Artificial Intelligence (XAI) suggests that a main aim of explainability approaches is to satisfy specific interests, goals, expectations, needs, and demands regarding artificial systems (we call these “stakeholders' desiderata”) in a variety of contexts. However, the literature on XAI is vast, spreads out across multiple largely disconnected disciplines, and it often remains unclear how explainability approaches are supposed to achieve the goal of satisfying stakeholders' desiderata. This paper discusses the main classes of stakeholders calling for explainability (...) of artificial systems and reviews their desiderata. We provide a model that explicitly spells out the main concepts and relations necessary to consider and investigate when evaluating, adjusting, choosing, and developing explainability approaches that aim to satisfy stakeholders' desiderata. This model can serve researchers from the variety of different disciplines involved in XAI as a common ground. It emphasizes where there is interdisciplinary potential in the evaluation and the development of explainability approaches. (shrink)

Many artificial intelligence (AI) systems currently used for decision-making are opaque, i.e., the internal factors that determine their decisions are not fully known to people due to the systems’ computational complexity. In response to this problem, several researchers have argued that human decision-making is equally opaque and since simplifying, reason-giving explanations (rather than exhaustive causal accounts) of a decision are typically viewed as sufficient in the human case, the same should hold for algorithmic decision-making. Here, I contend that this argument (...) overlooks that human decision-making is sometimes significantly more transparent and trustworthy than algorithmic decision-making. This is because when people explain their decisions by giving reasons for them, this frequently prompts those giving the reasons to govern or regulate themselves so as to think and act in ways that confirm their reason reports. AI explanation systems lack this self-regulative feature. Overlooking it when comparing algorithmic and human decision-making can result in underestimations of the transparency of human decision-making and in the development of explainable AI that may mislead people by activating generally warranted beliefs about the regulative dimension of reason-giving. (shrink)

In this report we focus on the definition of autonomous weapons systems (AWS). We provide a comparative analysis of existing official definitions of AWS as provided by States and international organisations, like ICRC and NATO. The analysis highlights that the definitions draw focus on different aspects of AWS and hence lead to different approaches to address the ethical and legal problems of these weapons systems. This approach is detrimental both in terms of fostering an understanding of AWS and in facilitating (...) agreement around conditions of deployment and regulations of their use and, indeed, whether AWS are to be used at all. We draw from the comparative analysis to identify essential aspects of AWS and then offer a definition that provides a value-neutral ground to address the relevant ethical and legal problems. In particular, we identify four key aspects—autonomy; adapting capabilities of AWS; human control; and purpose of use—as the essential factors to define AWS and which are key when considering the related ethical and legal implications. (shrink)

In this paper, in line with the general framework of value-sensitive design, we aim to operationalize the general concept of “Meaningful Human Control” in order to pave the way for its translation into more specific design requirements. In particular, we focus on the operationalization of the first of the two conditions investigated: the so-called ‘tracking’ condition. Our investigation is led in relation to one specific subcase of automated system: dual-mode driving systems. First, we connect and compare meaningful human control with (...) a concept of control very popular in engineering and traffic psychology, and we explain to what extent tracking resembles and differs from it. This will help clarifying the extent to which the idea of meaningful human control is connected to, but also goes beyond, current notions of control in engineering and psychology. Second, we take the systematic analysis of practical reasoning as traditionally presented in the philosophy of human action and we adapt it to offer a general framework where different types of reasons and agents are identified according to their relation to an automated system’s behaviour. This framework is meant to help explaining what reasons and what agents play a role in controlling a given system, thereby enabling policy makers to produce usable guidelines and engineers to design systems that properly respond to selected human reasons. In the final part, we discuss a practical example of how our framework could be employed in designing automated driving systems. (shrink)

The increasing demand for transparency in AI has recently come under scrutiny. The question is often posted in terms of “epistemic double standards”, and whether the standards for transparency in AI ought to be higher than, or equivalent to, our standards for ordinary human reasoners. I agree that the push for increased transparency in AI deserves closer examination, and that comparing these standards to our standards of transparency for other opaque systems is an appropriate starting point. I suggest that a (...) more fruitful exploration of this question will involve a different comparison class. We routinely treat judgments made by highly trained experts in specialized fields as fair or well-grounded even though—by the nature of expert/layperson division of epistemic labor—an expert will not be able to provide an explanation of the reasoning behind these judgments that makes sense to most other people. Regardless, laypeople are thought to be acting reasonably—and ethically—in deferring to the judgments of experts that concern their areas of specialization. I suggest that we reframe our question regarding the appropriate standards of transparency in AI as one that asks when, why, and to what degree it would be ethical to accept opacity in AI. I argue that our epistemic relation to certain opaque AI technology may be relevantly similar to the layperson’s epistemic relation to the expert in certain respects, such that the successful expert/layperson division of epistemic labor can serve as a blueprint for the ethical use of opaque AI. (shrink)

A classic objection to autonomous weapon systems (AWS) is that these could create so-called responsibility gaps, where it is unclear who should be held responsible in the event that an AWS were to violate some portion of the law of armed conflict (LOAC). However, those who raise this objection generally do so presenting it as a problem for AWS as a whole class of weapons. Yet there exists a rather wide range of systems that can be counted as “autonomous weapon (...) systems”, and so the objection is too broad. In this article I present a taxonomic approach to the objection, examining a number of systems that would count as AWS under the prevalent definitions provided by the United States Department of Defense and the International Committee of the Red Cross, and I show that for virtually all such systems there is a clear locus of responsibility which presents itself as soon as one focuses on specific systems, rather than general notions of AWS. In developing these points, I also suggest a method for dealing with near-future types of AWS which may be thought to create situations where responsibility gaps can still arise. The main purpose of the arguments is, however, not to show that responsibility gaps do not exist or can be closed where they do exist. Rather, it is to highlight that any arguments surrounding AWS must be made with reference to specific weapon platforms imbued with specific abilities, subject to specific limitations, and deployed to specific times and places for specific purposes. More succinctly, the arguments show that we cannot and should not aim to treat AWS as if all of these shared all morally relevant features, but instead on a case-by-case basis. Thus, we must contend with the realities of weapons development and deployment, and tailor our arguments and conclusions to those realities, and with an eye to what facts obtain for particular systems fulfilling particular combat roles. (shrink)

ABSTRACTAutonomous weapons systems pose many challenges in complex battlefield environments. Previous discussions of them have largely focused on technological or policy issues. In contrast, we focus here on the challenge of trust in an AWS. One type of human trust depends only on judgments about the predictability or reliability of the trustee, and so are suitable for all manner of artifacts. However, AWSs that are worthy of the descriptor “autonomous” will not exhibit the required strong predictability in the complex, changing (...) contexts of war. Instead, warfighters need to develop deeper, interpersonal trust that is grounded in understanding the values, beliefs, and dispositions of the AWS. Current acquisition, training, and deployment processes preclude the development of such trust, and so there are currently no routes for a warfighter to develop trust in an AWS. We thus survey three possible changes to current practices in order to facilitate the type of deep trust that is required for appropri... (shrink)

Due to advances in military technology, there has been an outpouring of research on what are known as autonomous weapon systems (AWS). However, it is common in this literature for arguments to be made without first making clear exactly what definitions one is employing, with the detrimental effect that authors may speak past one another or even miss the targets of their arguments. In this article I examine the U.S. Department of Defense and International Committee of the Red Cross definitions (...) of AWS, showing that these definitions are far broader than some recognize, and that they therefore classify a much larger set of weapons as AWS. I then show that these broader views of AWS have implications for what moral and legal rules we might argue should be applied to such systems. I conclude by arguing there is a greater need for precision and clarity within AWS debates, in order to ensure that researchers are discussing the same weapon systems (autonomous or otherwise) when they argue for or against particular points. The purpose of this article is not to defend any specific view of AWS, nor to further any general endorsement or objection to such systems, but rather to show the importance of argumentative clarity in this debate. (shrink)

A legal obligation to conduct weapons reviews is a means by which the international community can ensure that States assess whether the use of new types of weapons in armed conflict would raise humanitarian concerns. The use of artificial intelligence in weapon systems greatly complicates the process of conducting reviews, particularly where a weapon system is capable of continuing to ‘learn’ on its own after being deployed on the battlefield. This paper surveys current understandings of the weapons review challenges presented (...) by in situ learning and outlines some concepts which may inform appropriate changes to existing weapons review processes. The main arguments made are that: a distinction should be drawn between learning capabilities that are limited to optimising existing behaviour and capabilities that extend to acquiring new behaviours; and the extent of a weapon system’s learning capabilities should be used as the basis for defining a schedule of weapons reviews. (shrink)