I find this sort of thing very suggestive of some relationship to clinical/psychotic paranoia, and another recent HN thread made me think there might be fundamental challenges of deep learning systems that lead to similar problems:

https://news.ycombinator.com/item?id=17438733

When I looked at the example at the top of the paper, where Steve Carell is misidentified as Zooey Deschanel, I thought - I can envision the AI saying (anthropomorphically) yes, I know that looks like Steve overall, but just look at those eyes - that has to mean something. Kind of like a conspiracy theorist/paranoiac.