Deep learning techniques are growing in popularity within the field of artificial intelligence (AI). These approaches identify patterns in large scale datasets, and make classifications and predictions, which have been celebrated as more accurate than those of humans. But for a number of reasons, including nonlinear path from inputs to outputs, there is a dearth of theory that can explain why deep learning techniques work so well at pattern detection and prediction. Claims about “superhuman” accuracy and insight, paired with the inability to fully explain how these results are produced, form a discourse about AI that we call enchanted determinism. To analyze enchanted determinism, we situate it within a broader epistemological diagnosis of modernity: Max Weber’s theory of disenchantment. Deep learning occupies an ambiguous position in this framework. On one hand, it represents a complex form of technological calculation and prediction, phenomena Weber associated with disenchantment. On the other hand, both deep learning experts and observers deploy enchanted, magical discourses to describe these systems’ uninterpretable mechanisms and counter-intuitive behavior. The combination of predictive accuracy and mysterious or unexplainable properties results in myth-making about deep learning’s transcendent, superhuman capacities, especially when it is applied in social settings. We analyze how discourses of magical deep learning produce techno-optimism, drawing on case studies from game-playing, adversarial examples, and attempts to infer sexual orientation from facial images. Enchantment shields the creators of these systems from accountability while its deterministic, calculative power intensifies social processes of classification and control.

No abstract

Computer science tends to foreclose the reading of its texts by social science and humanities scholars – via code and scale, mathematics, black box opacities, secret or proprietary models. Yet, when computer science papers are read in order to better understand what machine learning means for societies, a form of reading is brought to bear that is not primarily about excavating the hidden meaning of a text or exposing underlying truths about science. Not strictly reading to make sense or to discern definitive meaning of computer science texts, reading is an engagement with the sense-making and meaning-making that takes place. We propose a strategy for reading computer science that is attentive to the act of reading itself, that stays close to the difficulty involved in all forms of reading, and that works with the text as already properly belonging to the ethico-politics that this difficulty engenders. Addressing a series of three “reading problems” – genre, readability, and meaning – we discuss machine learning textbooks and papers as sites where today's algorithmic models are actively giving accounts of their paradigmatic worldview. Much more than matters of technical definition or proof of concept, texts are sites where concepts are forged and contested. In our times, when the political application of AI and machine learning is so commonly geared to settle or predict difficult societal problems in advance, a reading strategy must open the gaps and difficulties of that which cannot be settled or resolved.

No abstract