The current state of AI image classification systems is limited and not reflective of the richness of human vision. A 2018 paper discusses the need for characterizing adversarial examples based on spatial consistency information for semantic segmentation.