In this talk, I discuss how the phenomenon of strange error bears differently on two distinct but overlapping questions. The first is about the possibility of using ML systems as sources of evidence for constructing human knowledge. The second is about the possibility of ascribing knowledge to ML systems themselves.