Nineteenth International Conference on Technology, Knowledge, and Society

Abstract

The standard interpretation of adversarial examples being used against image recognition algorithms, as an odd demonstration of erroneous behavior that embarrassingly shows the limitations of these developing technologies, has recently been called into question by the controversial “feature, not a bug” viewpoint advanced by Andrew Ilyas and others. Ilyas et al.’s research suggests a very different interpretation of the phenomenon, that rather than simply a mistake that needs to be overcome, it seems that at least in some cases, there is genuine information being conveyed that is “non-robust”–invisible to humans–and that these networks are responding to in some sense intelligibly. In my study, I discuss the further philosophical implications of this possibility for artificial intelligence, drawing a parallel between adversarial examples and optical illusions to which humans are susceptible. I argue that this suggests our definition of intelligent behavior is too narrowly defined around predictably “human” behavior, and that if we are seriously interested in defining and developing general artificial intelligence, we must make conceptual space and prepare ourselves for that general artificial intelligence to look much more alien than we have so far considered–what I call “strange artificial intelligence.”