Abstract

AbstractSimion and Kelp offer a prima facie very promising account of trustworthy AI. One benefit of the account is that it elegantly explains trustworthiness in the case of cancer diagnostic AIs, which involve the acquisition by the AI of a representational etiological function. In this brief note, I offer some reasons to think that their account cannot be extended — at least not straightforwardly — beyond such cases (i.e., to cases of AIs with non-representational etiological functions) without incurring the unwanted cost of overpredicting untrustworthiness.