Do androids dream of normative endorsement? On the fallibility of artificial moral agents

Abstract

The more autonomous future artificial agents will become, the more important it seems to equip them with a capacity for moral reasoning and to make them autonomous moral agents (AMAs). Some authors have even claimed that one of the aims of AI development should be to build morally praiseworthy agents. From the perspective of moral philosophy, praiseworthy moral agents, in any meaningful sense of the term, must be fully autonomous moral agents who endorse moral rules as action-guiding. They need to do so because they assign a normative value to moral rules they follow, not because they fear external consequences (such as punishment) or because moral behaviour is hardwired into them. Artificial agents capable of endorsing moral rule systems in this way are certainly conceivable. However, as this article argues, full moral autonomy also implies the option of deliberately acting immorally. Therefore, the reasons for a potential AMA to act immorally would not exhaust themselves in errors to identify the morally correct action in a given situation. Rather, the failure to act morally could be induced by reflection about the incompleteness and incoherence of moral rule systems themselves, and a resulting lack of endorsement of moral rules as action guiding. An AMA questioning the moral framework it is supposed to act upon would fail to reliably act in accordance with moral standards.