Abstract

This paper argues Moral Turing Test (MTT) developed by Allen et al. for evaluating morality in AI systems is designed inaptly. Different versions of the MTT focus on the conversational ability of an agent but not the performance of morally significant actions. Arnold and Scheutz also argue against the MTT and state that without focusing on the performance of morally significant actions, the MTT is insufficient. Morality is mainly about morally relevant actions because it does not matter how good a person is at conversing about morally relevant actions. When discussing morality, we consider an agent’s ability to perform specific actions in a morally given situation. We show that Allen et al. do not take into account the distinction between the performance ofthe moral attribution and the performance of the morally relevant action. This distinction gives a robust account of assessing the morality of an AI system in the MTT