Influencing machines: Trevor Paglen and Anthony Downey

Abstract How do you train an artificial intelligence (AI), or automated image processing model, to classify and recognize images? This question is central to Trevor Paglen’s Adversarially Evolved Hallucination series (2017–ongoing), a project that employs a generative adversarial network (GAN) to classify, identify and crucially, produce unique images. Paglen’s series demonstrates how images produced by AI image processing platforms—in this instance, a GAN—are, despite claims, never predictable or, indeed, accurate in their classifications. A significant indicator of this unreliability is evident in the potential for GANs, alongside other generative AI (GenAI) models, to hallucinate and erroneously classify images. Notwithstanding this systemic failing, automated image processing platforms remain central to classification tasks, including those associated with facial recognition and surveillance. They remain, for that reason, central to defining, if not pre-defining, how we perceive and look at the world through automated models of machine vision. Encouraged to see like machines, or at least take their classifications seriously and act upon them accordingly, we now inhabit a realm of perception defined by “machine realism”, if not algorithmic delusion. Enquiring into how we can better understand the degree to which AI encodes our perception of the world, it is this regimen of “machine realism” that Paglen and Downey explore throughout the following conversation: If AI models of image perception replace ocular-centric ways of seeing, they ask, do these apparatuses have the capacity to not only (pre)define but, in time, further estrange and alienate us from the world?