Atlas of AI

Atlas of AI refuses to begin with algorithms or the question of whether a machine can think. Kate Crawford moves to a larger geography: mines, logistics routes, data centers, warehouses, crowdwork platforms, training datasets, police agencies, and military systems. Her central claim is that AI is neither artificial nor independently intelligent. It depends on minerals and energy extracted from the earth, labor and data extracted from people, and decision-making power concentrated by a few companies and states. The “cloud” is not weightless; it has costs in land, water, electricity, waste, and bodies.

The chapters move through Earth, Labor, Data, Classification, Affect, and State like layers in a geological section. Hardware requires minerals such as lithium, global manufacturing, and energy-intensive infrastructure, while discarded devices move pollution into e-waste zones. This material perspective changes the unit of AI ethics. It is not enough to test an output for bias. A system must be evaluated across its life cycle—from extraction and training through deployment and disposal—and environmental damage must be traced to the communities farthest from the centers of technical profit.

Invisible labor is similarly layered beneath “automation.” Warehouse employees are managed by metrics and sensors; crowdworkers label images, clean datasets, and moderate content; every click or expression by a user may become training material. Such work is often low-paid, fragmented, and globally distributed, placing women, migrants, and other marginalized workers in positions of high risk and weak bargaining power. Crawford reverses the familiar story that machines replace people. AI often redistributes labor, hiding its most exhausting parts from the interface and the corporate account.

Data, too, are not natural resources waiting to be discovered. Collectors decide what enters a dataset, labelers compress continuous and ambiguous human differences into fixed categories, and models return those classifications as objective-looking predictions. Crawford follows the histories through which hierarchies of race, gender, ability, and normality enter technical systems. Her critique of affect recognition is especially useful: inferring inner emotion from facial movement packages disputed psychological theories as scalable products and gives schools, employers, and border agencies new tools to inspect bodies and doubt testimony.

When these systems enter government, the issue is larger than improving diversity on an engineering team. Predictive policing, benefit scrutiny, border control, and military targeting turn historical records into reasons for future intervention. Groups subjected to more surveillance generate more records, and those records become evidence of risk in a self-reinforcing loop. Feminist data justice must therefore move beyond “debiasing.” It asks whether a system should exist, who may refuse it, who can audit it, which judgments must not be delegated, and why public agencies place coercive authority inside opaque commercial infrastructure.

The atlas is both the book’s method and its limit. It connects sites and scales usually kept apart, allowing a reader to see the ore, assembly work, household data, platform, and waste behind a smart speaker. But every map selects a route. U.S. technology firms and state power form the main axis; regulation elsewhere, worker organizing, community technology, and alternative design receive less space. Crawford diagnoses extraction more fully than she specifies institutions capable of governing each supply-chain layer. Generative AI has since expanded model scale, but that change intensifies rather than erases her questions about resources, labor, and power.

For FemRes readers, the book expands “Is AI fair to women?” into a political-economic chain: whose land is mined, who labels and moderates, whose data are taken without meaningful consent, who is classified, and who can enforce a model’s output through state power? Any AI product can be traced backward along that chain. Read with Artificial Unintelligence, Algorithms of Oppression, Data Feminism, and work on digital labor. More just AI will require more than accurate models; it requires redistributing the authority to decide technology’s purposes, resources, and limits.