Human oversight has become a key mechanism for the governance of artificial intelligence (“AI”). Human overseers are supposed to increase the accuracy and safety of AI systems, uphold human values, and build trust in the technology. Empirical research suggests, however, that humans are not reliable in fulfilling their oversight tasks. They may be lacking in competence or be harmfully incentivised. This creates a challenge for human oversight to be effective. In addressing this challenge, this article aims to make three contributions. (...) First, it surveys the emerging laws of oversight, most importantly the European Union’s Artificial Intelligence Act (“AIA”). It will be shown that while the AIA is concerned with the competence of human overseers, it does not provide much guidance on how to achieve effective oversight and leaves oversight obligations for AI developers underdefined. Second, this article presents a novel taxonomy of human oversight roles, differentiated along whether human intervention is constitutive to, or corrective of a decision made or supported by an AI. The taxonomy allows to propose suggestions for improving effectiveness tailored to the type of oversight in question. Third, drawing on scholarship within democratic theory, this article formulates six normative principles which institutionalise distrust in human oversight of AI. The institutionalisation of distrust has historically been practised in democratic governance. Applied for the first time to AI governance, the principles anticipate the fallibility of human overseers and seek to mitigate them at the level of institutional design. They aim to directly increase the trustworthiness of human oversight and to indirectly inspire well-placed trust in AI governance. (shrink)

Feasting on a plethora of social media platforms, news aggregators, and online marketplaces, recommender systems (RSs) are spreading pervasively throughout our daily online activities. Over the years, a host of ethical issues have been associated with the diffusion of RSs and the tracking and monitoring of users’ data. Here, we focus on the impact RSs may have on personal autonomy as the most elusive among the often-cited sources of grievance and public outcry. On the grounds of a philosophically nuanced notion (...) of autonomy, we illustrate three specific reasons why RSs may limit or compromise it: the threat of manipulation and deception associated with RSs; the RSs’ power to reshape users’ personal identity; the impact of RSs on knowledge and critical thinking. In our view, however, notwithstanding these legitimate concerns, RSs may effectively help users to navigate an otherwise overwhelming landscape. Our perspective, therefore, is not to be intended as a bulwark to protect the _status quo_ but as an invitation to carefully weigh these aspects in the design of ethically oriented RSs. (shrink)

The value-ladenness of computer algorithms is typically framed around issues of epistemic risk. In this article, I examine a deeper sense of value-ladenness: algorithmic methods are not only themselves value-laden but also introduce value into how we reason about their domain of application. I call this domain distortion. In particular, using insights from jurisprudence, I show that the use of recidivism risk assessment algorithms presupposes legal formalism and blurs the distinction between liability assessment and sentencing, which distorts how the domain (...) of criminal punishment is conceived and provides a distinctive avenue for values to enter the legal process. (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)

Airline pilots are attributed ultimate responsibility and final authority over their aircraft to ensure the safety and well-being of all its occupants. Yet, with the advent of automation technologies, a dissonance has emerged in that pilots have lost their actual decision-making authority as well as their ability to act in an adequate fashion towards meeting their responsibilities when unexpected circumstances or emergencies occur. Across the literature in human factor studies, we show how automated algorithmic technologies have wrestled control away from (...) airline pilots. The inter-related ethical and pragmatic consequences resulting from such a dissonance are presented. This includes pilots serving as ‘legal sponges’, being caught in ‘moral crumple zones’, as well as real impacts on safety and well-being of passengers and crew. One potential avenue presented in the literature involves a redistribution of responsibility towards all agential entities within the cockpit system. We discuss the possible pitfalls to this approach, including the reinforcement of pilot skills erosion as well as technology’s own limitations in adaptively dealing with unexpected circumstances. We justify our position by showing how algorithmic technology’s representational frames mask or severely limit the pilots’ natural reflex to ‘practically cope’ and gain an ‘optimal grip’ with a given context/situation at hand. Here, we present the concept of relevant or meaningful control, which pilots, when allowed to develop their natural brain-body reflexes, gain across gradual mastery of the aircraft. Finally, we propose a socio-technical design approach in which pilots can re-appropriate full human meaningful control of the aircraft, thus allowing them to meet their attributed responsibilities. (shrink)

Accounting firms are reporting the use of Artificial Intelligence in their auditing and advisory functions, citing benefits such as time savings, faster data analysis, increased levels of accuracy, more in-depth insight into business processes, and enhanced client service. AI, an emerging technology that aims to mimic the cognitive skills and judgment of humans, promises competitive advantages to the adopter. As a result, all the Big 4 firms are reporting its use and their plans to continue with this innovation in areas (...) such as audit planning risk assessments, tests of transactions, analytics, and the preparation of audit work-papers, among other uses. As the uses and benefits of AI continue to emerge within the auditing profession, there is a gradual awakening to the fact that unintended consequences may also arise. Thus, we heed to the call of numerous researchers to not only explore the benefits of AI but also investigate the ethical implications of the use of this emerging technology. By combining two futuristic ethical frameworks, we forecast the ethical implications of the use of AI in auditing, given its inherent features, nature, and intended functions. We provide a conceptual analysis of the practical ethical and social issues surrounding AI, using past studies as well as our inferences based on the reported use of the technology by auditing firms. Beyond the exploration of these issues, we also discuss the responsibility for the policy and governance of emerging technology. (shrink)

Artificial intelligence-based clinical decision support systems have a potential to improve clinical practice, but they may have a negative impact on the physician-patient dialogue, because of the control problem. Physician-patient dialogue depends on human qualities such as compassion, trust, and empathy, which are shared by both parties. These qualities are necessary for the parties to reach a shared understanding -the merging of horizons- about clinical decisions. The patient attends the clinical encounter not only with a malfunctioning body, but also with (...) an ‘unhomelike’ experience of illness that is related to a world of values and meanings, a life-world. Making wise individual decisions in accordance with the patient’s life-world requires not only scientific analysis of causal relationships, but also listening with empathy to the patient’s concerns. For a decision to be made, clinical information should be interpreted considering the patient’s life-world. This side of clinical practice is not a job for computers, and they cannot be final decision-makers. On the other hand, in the control problem users blindly accept system output because of over-reliance, rather than evaluating it with their own judgement. This means over-reliant parties leave their place in the dialogue to the system. In this case, the dialogue may be disrupted and mutual trust may be lost. Therefore, it is necessary to design decision support systems to avoid the control problem and to limit their use when this is not possible, in order to protect the physician-patient dialogue. (shrink)

Technology could be used to improve morality but it could do so in different ways. Some technologies could augment and enhance moral behaviour externally by using external cues and signals to push and pull us towards morally appropriate behaviours. Other technologies could enhance moral behaviour internally by directly altering the way in which the brain captures and processes morally salient information or initiates moral action. The question is whether there is any reason to prefer one method over the other? In (...) this article, I argue that there is. Specifically, I argue that internal moral enhancement is likely to be preferable to external moral enhancement, when it comes to the legitimacy of political decision-making processes. In fact, I go further than this and argue that the increasingly dominant forms of external moral enhancement may already be posing a significant threat to political legitimacy, one that we should try to address. Consequently, research and development of internal moral enhancements should be prioritised as a political project. (shrink)

This paper examines what it takes for a state of knowledge-how to be extended (i.e. partly constituted by entities external to the organism) within an anti-intellectualist approach to knowledge-how. I begin by examining an account of extended knowledge-how developed by Carter, J. Adam, and Boleslaw Czarnecki. 2016 [“Extended Knowledge-How.” Erkenntnis 81 (2): 259–273], and argue that it fails to properly distinguish between cognitive outsourcing and extended knowing-how. I then introduce a solution to this problem which rests on the distribution of (...) tasks between agent and non-biological entity. On closer inspection, I show that this solution is ultimately unsatisfactory, though its failure is instructive as it illuminates the important role played by an agent’s skilled interaction with an external entity. Drawing on key anti-intellectualist ideas, as well as on insight from cognitive psychology, I propose an account according to which what ultimately matters for extending knowledge-how is whether a hybrid ability is self-regulated. In closing, I illuminate the practical value of extended knowledge-how vis-à-vis cognitive outsourcing. (shrink)

Can AI solve the ethical, moral, and political dilemmas of warfare? How is artificial intelligence (AI)-enabled warfare changing the way we think about the ethical-political dilemmas and practice of war? This article explores the key elements of the ethical, moral, and political dilemmas of human-machine interactions in modern digitized warfare. It provides a counterpoint to the argument that AI “rational” efficiency can simultaneously offer a viable solution to human psychological and biological fallibility in combat while retaining “meaningful” human control over (...) the war machine. This Panglossian assumption neglects the psychological features of human-machine interactions, the pace at which future AI-enabled conflict will be fought, and the complex and chaotic nature of modern war. The article expounds key psychological insights of human-machine interactions to elucidate how AI shapes our capacity to think about future warfare's political and ethical dilemmas. It argues that through the psychological process of human-machine integration, AI will not merely force-multiply existing advanced weaponry but will become de facto strategic actors in warfare – the “AI commander problem.”. (shrink)

Digital phenotyping is the term given to the capturing and use of user log data from health and wellbeing technologies used in apps and cloud-based services. This paper explores ethical issues in making use of digital phenotype data in the arena of digital health interventions. Products and services based on digital wellbeing technologies typically include mobile device apps as well as browser-based apps to a lesser extent, and can include telephony-based services, text-based chatbots, and voice-activated chatbots. Many of these digital (...) products and services are simultaneously available across many channels in order to maximize availability for users. Digital wellbeing technologies offer useful methods for real-time data capture of the interactions of users with the products and services. It is possible to design what data are recorded, how and where it may be stored, and, crucially, how it can be analyzed to reveal individual or collective usage patterns. The paper also examines digital phenotyping workflows, before enumerating the ethical concerns pertaining to different types of digital phenotype data, highlighting ethical considerations for collection, storage, and use of the data. A case study of a digital health app is used to illustrate the ethical issues. The case study explores the issues from a perspective of data prospecting and subsequent machine learning. The ethical use of machine learning and artificial intelligence on digital phenotype data and the broader issues in democratizing machine learning and artificial intelligence for digital phenotype data are then explored in detail. (shrink)

The rapid adoption and implementation of artificial intelligence in medicine creates an ontologically distinct situation from prior care models. There are both potential advantages and disadvantages with such technology in advancing the interests of patients, with resultant ontological and epistemic concerns for physicians and patients relating to the instatiation of AI as a dependent, semi- or fully-autonomous agent in the encounter. The concept of libertarian paternalism potentially exercised by AI (and those who control it) has created challenges to conventional assessments (...) of patient and physician autonomy. The unclear legal relationship between AI and its users cannot be settled presently, an progress in AI and its implementation in patient care will necessitate an iterative discourse to preserve humanitarian concerns in future models of care. This paper proposes that physicians should neither uncritically accept nor unreasonably resist developments in AI but must actively engage and contribute to the discourse, since AI will affect their roles and the nature of their work. One’s moral imaginative capacity must be engaged in the questions of beneficence, autonomy, and justice of AI and whether its integration in healthcare has the potential to augment or interfere with the ends of medical practice. (shrink)

Legislation and ethical guidelines around the globe call for effective human oversight of AI-based systems in high-risk contexts – that is oversight that reliably reduces the risks otherwise associated with the use of AI-based systems. Such risks may relate to the imperfect accuracy of systems (e.g., inaccurate classifications) or to ethical concerns (e.g., unfairness of outputs). Given the significant role that human oversight is expected to play in the operation of AI-based systems, it is crucial to better understand the conditions (...) for effective human oversight. We argue that the reliable detection of errors (as an umbrella term for inaccuracies and unfairness) is crucial for effective human oversight. We then propose that Signal Detection Theory (SDT) offers a promising framework for better understanding what affects people’s sensitivity (i.e., how well they are able to detect errors) and response bias (i.e., the tendency to report errors given a perceived evidence of an error) in detecting errors. Whereas an SDT perspective on the detection of inaccuracies is straightforward, we demonstrate its broader applicability by detailing the specifics for an SDT perspective on unfairness detection, including the need to choose a standard for (un)fairness. Additionally, we illustrate that an SDT perspective helps to better understand the conditions for effective error detection by showing examples of task-, system-, and person-related factors that may affect the sensitivity and response bias of humans tasked with detecting unfairness associated with the use of AI-based systems. Finally, we discuss future research directions for an SDT perspective on error detection. (shrink)

We argue that cognitive models can provide a common ground between human users and deep reinforcement learning (Deep RL) algorithms for purposes of explainable artificial intelligence (AI). Casting both the human and learner as cognitive models provides common mechanisms to compare and understand their underlying decision-making processes. This common grounding allows us to identify divergences and explain the learner's behavior in human understandable terms. We present novel salience techniques that highlight the most relevant features in each model's decision-making, as well (...) as examples of this technique in common training environments such as Starcraft II and an OpenAI gridworld. (shrink)

The danger of human operators devolving responsibility to machines and failing to detect cases where they fail has been recognised for many years by industrial psychologists and engineers studying the human operators of complex machines. We call it “the control problem”, understood as the tendency of the human within a human–machine control loop to become complacent, over-reliant or unduly diffident when faced with the outputs of a reliable autonomous system. While the control problem has been investigated for some time, up (...) to this point its manifestation in machine learning contexts has not received serious attention. This paper aims to fill that gap. We argue that, except in certain special circumstances, algorithmic decision tools should not be used in high-stakes or safety-critical decisions unless the systems concerned are significantly “better than human” in the relevant domain or subdomain of decision-making. More concretely, we recommend three strategies to address the control problem, the most promising of which involves a complementary coupling between highly proficient algorithmic tools and human agents working alongside one another. We also identify six key principles which all such human–machine systems should reflect in their design. These can serve as a framework both for assessing the viability of any such human–machine system as well as guiding the design and implementation of such systems generally. (shrink)

Here we present a systematic plan to the experimental study of test–retest reliability in the multitasking domain, adopting the multitrait-multimethod approach to evaluate the psychometric properties of performance in Düker-type speeded multiple-act mental arithmetic. These form of tasks capacitate the experimental analysis of integrated multi-step processing by combining multiple mental operations in flexible ways in the service of the overarching goal of completing the task. A particular focus was on scoring methodology, particularly measures of response speed variability. To this end, (...) we present data of two experiments with regard to test–retest reliability, between-measures correlational structure, and stability. Finally, we compared participants with high versus low performance variability to assess ability-related differences in measurement precision, which is especially relevant in the applied fields of clinical neuropsychology. The participants performed two classic integrated multi-act arithmetic tasks, combining addition and verification and addition and comparison. The results revealed excellent test–retest reliability for the standard and the variability measures. The analysis of between-measures correlational structure revealed the typical pattern of convergent and discriminant relationships, and also, that absolute response speed variability was highly correlated with average speed, indicating that these measures mainly deliver redundant information. In contrast, speed-adjusted variability revealed discriminant validity being correlated to a much lesser degree with average speed, indicating that this measure delivers additional information not already provided by the speed measure. Furthermore, speed-adjusted variability was virtually unaffected by test–retest practice, which makes this measure interesting in situations with repeated testing. (shrink)

Process reliabilist accounts claim that a belief is justified when it is the result of a reliable belief-forming process. Yet over what range of possible token processes is this reliability calculated? I argue against the idea that _all_ possible token processes (in the actual world, or some other subset of possible worlds) are to be considered using the case of a user acquiring beliefs based on the output of an AI system, which is typically reliable for a substantial local range (...) but unreliable when all possible inputs are considered. I show that existing solutions to the generality problem imply that these cases cannot be solved by a more fine-grained typing of the belief-forming process. Instead, I suggest that reliability is evaluated over a range restricted by the content of the actual belief and by the similarity of the input to the actual input. (shrink)