The Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence Kate Crawford, The Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Pp. 1, 327. ISBN 9780300209570, 9780300264630. Review by Stella J. Fritzell, Bryn Mawr College. sfritzell@brynmawr.edu
This volume sets out to explore how artificial intelligence is produced, winding its way through the various historical, cultural, economic, and political forces that inform and shape this process. Crawford frames her discussion not as a technical manual about code, algorithms, computer vision, language processing, and reinforcement learning methods, but rather as “an expanded view of artificial intelligence as an extractive industry,” which exploits energy, mineral resources, cheap labor, and data at scale (15). In this regard, The Atlas of AI is thoroughly illuminating. With this book, Crawford convincingly argues that AI is neither artificial nor intelligent, but rather stands as a registry of power—a system built upon natural- and human resources, and dependent on much larger socio-political structures and vast amounts of economic capital. The Atlas of AI consists of six chapters with an introduction, conclusion, and final coda, acknowledgements, and twenty-four pages of notes, as well as a robust bibliography and index. A table of contents is printed at the end of this review. The central question of how intelligence is made and accurately measured is posed in the introduction, which additionally lays out the core disputes that underly the development of artificial intelligence and thus are also present throughout the discussions of this book—whether humans and human intelligence may be accurately and fairly understood as information processing systems, and thus whether it is possible to recreate such a processing system in the guise of a machine. The volume argue that such is not the case. A so-called artificial “intelligence” can only replicate human processes in as much as it relies on these processes for its creation and upkeep. These human processes, from minute to sweeping in scale, are detailed in their varying forms in chapters two through six. Chapter Two demonstrates how the increasing commodification of labor has paralleled the development of artificial intelligence systems for worker surveillance and assessment, and lays out the exploitative labor practices, such as on-demand crowdwork, upon which AI systems rely. The transformation of data from personal material into impersonal infrastructure is the subject of Chapter Three, which discusses the ethical, methodological, and epistemological concerns that are (or ought to be) part of this process of data collection and use. Chapter Four explores how the ordering schemata emergent in AI training datasets naturalize rigid hierarchies and magnify extant social inequalities. Chapter Five picks up on the issues of facial recognition introduced in previous chapters and turns particularly to the question of affect detection. Crawford points out that while “there is no reliable evidence that you can accurately predict someone’s emotional state from their face” (120), the systems developed on such false assumptions are nevertheless highly effective at introducing recognizable patterns of behavior among individuals. As the final chapter, Chapter Six discusses how state and corporate entities have employed AI in surveillance and intelligence efforts. The Conclusion naturally returns to the question of AI as an autonomous “intelligence” and reiterates how the stakes of AI are “not the technocratic imaginaries of artificiality, abstraction, and automation,” but rather global infrastructures of extraction and power (218). In her brief Coda, Crawford turns to the ongoing space-race funded by billionaire corporate leaders and discusses how this embodies the same systems of inequality and extractive mindsets discussed throughout the volume. Human actors and actions are uncovered in Chapter One through the discussion of their environmental effects. Beginning with the lithium deposits at Silver Peak in Nevada’s Clayton Valley, moving to the Tesla Gigafactory, mining operations in Mongolia, the physical infrastructure of the cloud, and finally to the global shipping industry, Chapter One seeks to track the logics of extraction across (modern) global time and space. Crawford defines these logics that lay implicit in large-scale computation as “a constant drawdown of minerals, water, and fossil fuels, undergirded by the violence of wars, pollution, extinction, and depletion” (28). Informed by Jussi Parikka’s A Geology of Media, Crawford argues in this chapter that media and technology are best conceptualized as geological processes. The roots of AI and the ambiguously-named Cloud rest in the composites, minerals, and crude oil that are extracted from the Earth. The matter of environmental and sociological devastation wrought by mining and related industries has long been an open secret. At least 23 rare-earth minerals, for example, have been named by the U.S. Geological Survey as a high “supply risk”. These minerals exist in limited natural quantities, are incredibly difficult to extract from the Earth’s crust, and produce substantial amounts of waste toxins through the refinement processes that render them useable. Yet because the electronic, optical, and magnetic uses of these minerals, such as Lithium, remain unmatched, Crawford demonstrates that the demand for these resources only continues to grow, even as the reserves themselves dwindle. This, in turn, has given rise to the local and geopolitical violence that often accompanies the extraction of such resources. Should their supply be completely exhausted, entire industries would cease to function, totally upending daily life. Connected to this are the high energy costs involved both in the extraction and transportation of such minerals, and in the maintenance of computation infrastructures that support regular—or what are presently considered to be regular—uses of the internet and the development of tools like AI. The amount of annual carbon emissions for the tech sector, for example, has in recent years exceeded that of the global airline industry, even as tech companies continue to publicize their environmental and sustainability initiatives which position AI as a problem-solving tool. To quote Crawford:
The corporate imaginaries of AI fail to depict the lasting costs and long histories of the materials needed to build computation infrastructures or the energy required to power them. The rapid growth of cloud-based computation, portrayed as environmentally friendly, has paradoxically driven an expansion of the frontiers of resource extraction. (47)
Crawford describes this failure to account for the negative effects of mining and resource extraction as part of a “strategic amnesia” that goes hand-in-hand with technological advancement (26). Although it is not characterized as such, a reader may be tempted to think of this amnesia as malicious, a deliberate effort to ignore and obscure the ugly sides of commercial development intended to make lives more comfortable. And while the modern appetite for near-constant technological evolution has been accompanied by a proportionate growth in the scale and speed demanded of extractive industries, this associated amnesia is much more ancient. The environmental effects of mining and related pyrotechnic industries in the ancient Mediterranean, for example, were significant. Yet ancient commentators generally appear to ignore these impacts, possibly because the effect of such industries on the environment at the time was more gradual, and thus difficult to observe.[1] Unfortunately, the repercussions of modern computation for the ecological landscape are neither minute nor gradual in scope. I would suggest, that like the ancient processes of metallurgy and ceramic production, the tech sector is considered “constructive”, in that it produces something which is not itself environmentally detrimental. The difficulty of reconciling this cognitive dissonance between constructive and destructive sides of production certainly has something to do with the selective amnesia described by Crawford, and this chapter aids in bridging this divide. In in her Conclusion, Crawford introduces the phrase “enchanted determinism” to express how artificial intelligence is typically imagined as an object that simultaneously exists beyond the known world and exerts influence over daily life (213). This way of viewing AI encompasses both optimistic readings of tech utopianism and pessimistic predictions of a dystopian future. While this book is written for those curious about the non-technical aspects of AI development, it will appeal most directly to audiences already inclined to view AI in a negative light. Those more favorably minded towards the rapid technological advancements of the 21st century are less likely to look deeply into the processes of AI production, and thus less likely to seek out this volume. While The Atlas of AI does an excellent job of laying bare the elements of technical development frequently forgotten as a result of the aforementioned “strategic amnesia”, for those already content with this mirage there is little impetus to open the cover. The question then stands, how can such individuals be made aware of the systems of AI development, and how can the practices of tech companies best be audited if public awareness remains unchanged and unchallenged? Those who do take up this book will be met with clear and engaging language, and an appealing trajectory of discussion. Crawford provides her readers with a well-planned journey, laying out concrete examples to demonstrate how artificial intelligence is constructed and trained, from abstract conception to material reality.
Contents: Introduction (1) 1. Earth (23) 2. Labor (53) 3. Data (89) 4. Classification (123) 5. Affect (151) 6. State (181) Conclusion: Power (211) Coda: Space (229) Acknowledgements (239) Notes (245) Bibliography (269) Index (315)
Notes: [1] Cf. Wertime, Theodore A. 1983. “The Furnace versus the Goat: The Pyrotechnologic Industries and Mediterranean Deforestation in Antiquity.” Journal of Field Archaeology, 10.4: 445-452.