Toward a phenomenology of mystification: The ambivalence of magic among AI technologists

By leveraging machine learning, computer scientists and engineers have developed generative AI models so opaque that experimenting on them is often the only means to study their inner workings. In our ethnographic engagements with such technologists, we have frequently witnessed them contemplate model outputs with rapt mystification. Although accusations that AI technologists deliberately mystify investors, regulators, and consumers are certainly valid, we argue that these critiques also obscure the mystification researchers themselves experience. Taking the anthropology of magic as a point of departure, we focus on those experiences to theorize an anthropological phenomenology of mystification. Discussing the awe experienced by AI researchers, we explore how the value of mystification may be not only a byproduct of building constitutively opaque systems, but also an emergent ambition that increasingly animates the design process. Through a contrast with systems of divination, we argue that tech industry scientists experiment on AI systems not only to overcome opacity, but also to stabilize it and make it available as a resource for future interpretative acts. Far from demystifying generative models, interpretability experiments may reproduce the very sense of enchanting metaphysical excess they nominally seek to contain.