Humans can predict how machines (mis)classify adversarial images

We question how AI could label an image of a dog as a pineapple. But research shows there's a logic to these errors that humans can intuitively understand.