Abstract
The ethics of robots and artificial intelligence (AI) typically centers on “giving ethics” to as-yet imaginary AI with human-levels of autonomy in order to protect us from their potentially destructive power. It is often assumed that to do that, we should program AI with the true moral theory (whatever that might be), much as we teach morality to our children. This paper argues that the focus on AI with human-level autonomy is misguided. The robots and AI that we have now and in the near future are “semi-autonomous” in that their ability to make choices and to act is limited across a number of dimensions. Further, it may be morally problematic to create AI with human-level autonomy, even if it becomes possible. As such, any useful approach to AI ethics should begin with a theory of giving ethics to semi-autonomous agents (SAAs). In this paper, we work toward such a theory by evaluating our obligations to and for “natural” SAAs, including nonhuman animals and humans with developing and diminished capacities. Drawing on research in neuroscience, bioethics, and philosophy, we identify the ways in which AI semi-autonomy differs from semi-autonomy in humans and nonhuman animals. We conclude on the basis of these comparisons that when giving ethics to SAAs, we should focus on principles and restrictions that protect human interests, but that we can only permissibly maintain this approach so long as we do not aim at developing technology with human-level autonomy.