As Artificial Intelligence (AI) systems shift to interact with new domains and populations, so does AI ethics: a relatively nascent subdiscipline that frequently concerns itself with questions of “fairness” and “accountability.” This fairness-centred approach has been criticized for (amongst other things) lacking the ability to address discursive, rather than distributional, injustices. In this paper I simultaneously validate these concerns, and work to correct the relative silence of both conventional and critical AI ethicists around disability, by exploring the narratives deployed by AI researchers in discussing and designing systems around autism. Demonstrating that these narratives frequently perpetuate a dangerously dehumanizing model of autistic people, I explore the material consequences this might have. More importantly, I highlight the ways in which discursive harms—particularly discursive harms around dehumanization—are not simply inadequately handled by conventional AI ethics approaches, but actively invisible to them. I urge AI ethicists to critically and immediately begin grappling with the likely consequences of an approach to ethics which focuses on personhood and agency, in a world in which many populations are treated as having neither. I suggest that this issue requires a substantial revisiting of the underlying premises of AI ethics, and point to some possible directions in which researchers and practitioners might look for inspiration.