Limitations and risks of machine ethics

Many authors have proposed constraining the behaviour of intelligent systems with " machine ethics " to ensure positive social outcomes from the development of such systems. This article critically analyses the prospects for machine ethics, identifying several inherent limitations. While machine ethics may increase the probability of ethical behaviour in some situations, it cannot guarantee it due to the nature of ethics, the computational limitations of computational agents, and the complexity of the world. Additionally, machine ethics, even if it were to be " solved " at a technical level, would be insufficient to ensure positive social outcomes from intelligent systems. Introduction In response to the increasing sophistication and social impact of artificial intelligence-based technologies, some authors have proposed that software agents and robots be imbued with explicit ethical principles to govern their behaviour (Allen & Wallach, 2010). This area of research, known as " machine ethics, " incorporates insights from normative ethics, AI, and other disciplines and seeks to instantiate models of ethical reasoning in machines in order to ensure ethical machine behaviour as well as to explore the nature and computability of human ethics. In addition, other authors concerned with the risks posed by future greater-than-human intelligence machines have characterized machine ethics as one among multiple strategies for ensuring positive social outcomes from the creation of such machines. Yudkowsky (2001), for example, argues that failing to develop a sufficiently well-formulated computational instantiation of morality will result in catastrophic outcomes from developing superintelligent machines, since the instrumental goals of self-improving systems seeking to maximize some arbitrary utility function will at some point come into conflict with human interests. What these literatures (referred to here as " traditional machine ethics " on the one hand, and " artificial general intelligence (AGI) ethics " on the other, though the two have cross-pollinated to some extent) have in common is the apparent belief that the ethical behaviour " problem " can be " solved " to some extent through philosophical and technological innovation. While this article does not seek to undermine the desirability of such a solution, in principle, it points to a number of reasons to believe that such a project will necessarily fail to guarantee ethical behaviour of a given AI system across all possible domains. The intrinsic imperfectability of machine ethics has been suggested by several authors, such as Allen et al. (2000); the present paper seeks to synthesize, …