Kate Crawford. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. New Haven: Yale University Press, 2021. 336 pp.
Review by Lauren M. E. Goodlad
25 April 2022
Depending on your familiarity with artificial intelligence (AI), Kate Crawford’s Atlas of AI will either introduce you to a much-hyped revolution in computer technology or reprise an important set of critiques in an absorbing way. Crawford, an interdisciplinary anthropologist whose AI studies traverse research and activism, divides her book into thematic chapters such as “Labor,” “Classification,” “Affect,” and “Power.” While these discussions are thickly annotated, compellingly historicized, and often brilliantly synthetic, the author also undertakes ethnographic excursions to frame her chapters with vivid insights into fulfillment centers, government archives, museums, and mines. The result is a book that will beguile newcomers as they tour technology’s deep hold on the twenty-first century, while offering seasoned readers a cogent compendium for research and teaching.
As some readers will be aware, the dominant AI of our time has little to do with the human-like androids of fiction and film. To make matters more confusing, mainstream discussions are laced with triumphalist declarations of a fourth industrial revolution, mystifying talk of “Turing tests,” a specialist jargon of “neural networks” (dating back to the Cold War), and breathless forewarnings of robot takeover and mass unemployment. In the real world, the most powerful AIs are modes of data-driven machine learning: software programs that mobilize arsenals of data and computational power to produce statistical models that proffer useful (or at least lucrative) predictions about people or the world. The result is a multipurpose technology that propagates mass surveillance, transforms patterns from the past into projections of the future, and concentrates decision-making in the hands of those who design, implement, and profit from AI. One of Crawford’s central points is that AI is actually neither artificial (since it relies on human-generated data, programming, and assistance) nor intelligent (in the human sense of generalizing from experience, imagining hypotheticals, or apprehending the world through common sense and causal reasoning). Moreover, given the “capital required to build” it, and “the ways of seeing” which the technology activates, AI, above all, is “a registry of power” (p. 8). As it remaps, disrupts, intervenes, and codifies, AI takes the form of "politics by other means" (p. 20).
Crawford’s method is nicely illustrated by “Earth,” the first and most atlas-like of her chapters. Her first-person narration lands the reader in San Francisco by way of preparing us for Silicon Valley’s pride of place in a “planetary supersystem” interweaved through a “logics of extraction” (p. 28). Though we read about Chinese rare minerals and Malaysian latex as well, Crawford personally escorts us to Silver Peak, an old mining town in Nevada, where an “underground lake of lithium” now proffers “the stuff” of which AI “is made” (pp. 24–25). We learn that “the lifecycle of an AI system” is a matter of far-flung supply chains that drain energy at every step (p. 32). Such third-party networks militate toward the public’s ignorance of environmental and human harms, while providing tech companies with “plausible deniability for any exploitative practices that drive their profits” (p. 35). Perhaps the epitome of this delusive regime is “the cloud,” an Orwellian misnomer that belies the fact that data centers consume energy and water at the scale of large nation-states. In a recurrent pattern of ethnographic framing, the chapter closes where it began: the Nevada lithium pool, “extracted for batteries that are destined for landfill,” will one day be depleted like the abandoned silver mines of yore (p. 51).
“Data,” Crawford’s third chapter, opens with a series of mugshots from “Special Database 32,” maintained by the National Institute of Standards of Technology (an arbiter of scientific measurement since 1901). As Crawford shows, the AI traffic in biometric data connects this millennial tech to the eugenicist pseudo-sciences of the late nineteenth century. In a more contemporary vein, “datafication” is the process through which information about people becomes part of an “aggregate mass” that drives a “broader system,” the ostensible justification for which is improved “technological performance” (pp. 114, 93–94). That is, as data becomes ever more necessary to AI’s statistical engines for “learning”—and as Big Tech companies push a logic of bigger-is-best which serves their interests—a research field that was never equipped to examine ethical quandaries has become willfully blind. The business practices of Google and Facebook—platforms that turn user information into climbing share price—encourage researchers to regard the “ethical, political, and epistemological problems” of mining data without care or consent as just another externality for someone else to deal with (p. 111). Their task, on this view, is to chase after progress or face irrelevance. But as machine learning increasingly includes tools for policing, sentencing, benefits, lending, healthcare, hiring, and admissions, the social fallout from building inaccurate and prejudicial AI becomes ever more grave. That said, technologists are among those who have begun to take notice: in concert with books like Crawford’s, a new wave of socially conscious research and policy centers has emerged (including the AI Now Institute, the Algorithmic Justice League, and the Distributed AI Research Institute).
Crawford’s understanding of machine learning technologies is admirably sound: but a fuller discussion of AI function is peripheral to her political and environmental analyses. Readers keen to look under the hood will get a stronger technical introduction from Cathy O’Neil’s Weapons of Math Destruction (2016), Meredith Broussard’s Artificial Unintelligence (2018), or (for an even deeper dive) Gary Marcus and Ernest Davis’s Rebooting AI (2019). And while Crawford is duly attentive to the pervasive biases of AI systems, those interested in race and gender should not miss Safiya Noble’s Algorithms of Oppression (2018) or Ruha Benjamin’s Race after Technology (2019). As an encompassing and trenchant survey of what informed twenty-first century citizens should know, Atlas of AI is an invaluable entry point, a guide to pressing interdisciplinary questions, and a call to arms.