ABSTRACT
This chapter distinguishes among four different kinds of responsibility gaps and briefly compares them to four kinds recently identified by Filippo Santoni de Sio and Giulio Mecacci. A responsibility gap occurs when some event/outcome seems to be such that it would be fitting to hold somebody responsible for it, but where there is no person or persons who could fittingly be held responsible. The four responsibility gaps discussed relate to the ideas of forward-looking and backward-looking responsibility, and positive and negative responsibility. This chapter focuses specifically on so-called forward-looking positive responsibility gaps and relates such gaps to the so-called value alignment problem in artificial intelligence (AI) ethics. This is the problem of making sure that advanced AI aligns with human values, interests, or aims, so that risks related to advanced AI are mitigated. Lastly, this chapter compares two attempts, one by Nyholm and one by Santoni de Sio and Mecacci, to fill responsibility gaps. Both solutions can be viewed as attempts to specify what it is to have meaningful human control over AI. Both attempted solutions, this chapter argues, face important problems. In particular, these theories both work best when we think of them as idealized models for how to fill responsibility gaps. But both theories become more problematic when it comes to the issue of how to actually put them to work in practice in relation to real-world risks (e.g., risks related to advanced AI).
