Abstract The use of autonomous artificial intelligence systems in sensitive domains requires them to decide and act ethically. The field of machine ethics is dedicated to developing such artificial moral agents. This survey classifies recently published artificial moral agent implementations to identify trends and to evaluate progress on existing challenges within the field. To this end, we deploy a taxonomy composed of three dimensions: (1) ethical implementation aspects, (2) non-technical implementation aspects, and (3) technical implementation aspects. We identify a trend towards systems which are configurable with regard to the ethical theory implemented, as well as a trend towards learning-based approaches. Not using benchmarks to evaluate an agent’s performance and the lack of code availability remain open challenges. We argue that the absence of moral patient involvement is an equally big issue: those who are affected by a system’s decisions are rarely involved in specifying the ethical principles or rules according to which the system acts. The supplementary material provides short summaries of all surveyed papers.
Krah et al. (Sun,) studied this question.