Joris Graff

With the advent of explainable artificial intelligence (XAI) to explain the outputs of black-box machine learning models, the question arises how such explanations should be conceptualised. Specifically, assuming that XAI methods provide explanations, what types of explanations are these? Although this question is usually left implicit, informal discussion of XAI methods often suggests that XAI provides reason explanations that show a machine learning model’s reasons for its classifications. This suggestion has recently been explicitly defended by Zerilli and Baum et al. This paper argues that this idea is mistaken. Neither of the two main families of views of reason explanations—the causal/dispositional family, on the one hand, and the interpretative family, on the other—can be applied to XAI. On the first type of view, it is unclear how the causal or dispositional states of black-box models could ground the normative attitudes characteristic of motivating reasons. On the second type of view, a reason explanation should show a decision in a positive light from the decision-maker’s perspective. This means that the provider of a reason explanation should be able to (accurately) answer normative why-questions about the decision-maker. It also requires the assumption that the decision-maker has recognisable concerns and beliefs. Black-box models, and XAI methods designed to explain them, do not meet these conditions. It is concluded that other types of explanations, such as causal explanations, are more appropriate to conceptualise XAI outputs, and that a reason-based conceptualisation risks creating false expectations and possibly over-reliance in users.

The algorithmic curation of social media content and advertisements has a significant impact on many people’s behaviour. This influence is leading to increasing moral unease, calling for ethical reflection. A common analysis conceptualises problematic online influence as a form of manipulation. This paper argues that this analysis is flawed, since standard accounts of manipulation presuppose a manipulative intention or attitude, which cannot be identified in many cases of algorithmic curation. Instead, it is argued that an account of problematic online influence should focus on the risks posed to social media users’ personal autonomy. An outline of such an account is sketched by identifying two conditions for personal autonomy and arguing that algorithmic curation is likely to undermine both conditions. First, personal autonomy requires the capacity to act for normative reasons. However, social media algorithms often make users act for nonreasons and may erode capacities needed to recognise normative reasons. Second, personal autonomy requires that the reasons for which someone acts cohere with that person’s commitments and life plans. However, social media algorithms invite users to act for reasons isolated from their broader commitments. The conclusion is that social media platforms create wrongful risks to their users’ autonomy.

Machine ethics is the field that strives to develop ‘artificial moral agents’ (AMAs), artificial systems that can autonomously make moral decisions. Some authors have questioned the feasibility of machine ethics, by questioning whether artificial systems can possess moral competence, or the capacity to reach morally right decisions in various situations. This paper explores this question by drawing on the work of several moral philosophers (McDowell, Wiggins, Hampshire, and Nussbaum) who have characterised moral competence in a manner inspired by Aristotle. Although disparate in many ways, these philosophers all emphasise what may be called ‘moral sensitivity’ as a precondition for moral competence. Moral sensitivity is the uncodified, practical skill to recognise, in a range of situations, which features of the situations are morally relevant, and how they are relevant. This paper argues that the main types of AMAs currently proposed are incapable of full moral sensitivity. First, top-down AMAs that proceed from fixed rule-sets are too rigid to respond appropriately to the wide range of qualitatively unique factors that moral sensitivity gives access to. Second, bottom-up AMAs that learn moral behaviour from examples are at risk of generalising from these examples in undesirable ways, as they lack embedding in what Wittgenstein calls a ‘form of life’, which allows humans to appropriately learn from moral examples. The paper concludes that AMAs are unlikely to possess full moral competence, but closes by suggesting that they may still be feasible in restricted domains of public morality, where moral sensitivity plays a smaller role.