Atlas of AI

Kate Crawford argues that artificial intelligence is not an abstract, neutral technology but a material system of extraction and power, built from natural resources, human labor, and data, and designed to serve dominant interests.

She opens with the parable of Clever Hans, a German horse that appeared to solve math problems in the early 1900s but was actually responding to unconscious physical cues from his questioners. The story exposes two myths central to AI: that nonhuman systems can function as analogues for human minds, and that intelligence exists independently of social, cultural, and political forces. Crawford contends that AI is neither artificial nor intelligent in any meaningful sense but is embodied and material, dependent on mineral resources, fuel, infrastructure, logistics, histories, and classification systems. She introduces the atlas as a structuring metaphor, drawing on historian Lorraine Daston's observation that atlases train the eye to see in particular ways. Against the AI industry's ambition to map the world on its own terms, Crawford offers a grounded account organized around six sites: mineral extraction, human labor, data harvesting, classification systems, emotion detection, and state power.

The first chapter begins at the lithium mines of Silver Peak, Nevada. Crawford draws a parallel between San Francisco's nineteenth-century enrichment through gold and silver mining and its current enrichment through the tech sector, observing that mining has only been profitable because it externalizes its true costs: environmental damage, illness and death of miners, and displacement of communities. She details the lithium supply chain and describes the human toll of extraction, from decades of armed conflict in the Democratic Republic of the Congo financed by mining profits to the toxic artificial lake in Baotou, Inner Mongolia, holding over 180 million tons of waste from rare earth mineral processing. Crawford challenges the myth of clean technology, citing research showing that training a single natural language processing model produced more than 660,000 pounds of carbon dioxide emissions, and invokes historian Lewis Mumford's concept of the megamachine to characterize AI as a system dependent on opaque global infrastructures.

The second chapter moves to an Amazon fulfillment center in Robbinsville, New Jersey, where workers wear support bandages and have access to vending machines stocked with painkillers. Crawford argues that the central question is not whether robots will replace humans but how humans are increasingly treated like robots through surveillance, algorithmic assessment, and the modulation of time. She traces the prehistory of workplace AI through economist Adam Smith's division of labor and political theorist Karl Marx's observation that automation turns workers into machine appendages. She describes engineer Samuel Bentham's inspection house, a 1780s Russian workplace surveillance mechanism that inspired his brother Jeremy Bentham's panopticon prison. Crawford introduces "ghost work," scholars Mary Gray and Sid Suri's term for hidden crowdworkers who maintain AI systems for pennies per task, and "fauxtomation," writer Astra Taylor's term for systems that appear automated but relocate labor to underpaid workers. She reports on Amazon warehouse worker organizing in Minnesota, where workers demanded changes to "the rate," the productivity metric Amazon refused to negotiate because it constituted the company's business model.

The third chapter examines data extraction, opening with a federal database of thousands of mug shot photographs now used as a benchmark for facial recognition systems. Crawford traces the history of data demand from science administrator Vannevar Bush's 1945 vision of machines with "enormous appetites" through IBM's shift from linguistic to statistical methods in speech recognition. She describes the creation of ImageNet beginning in 2006, when Stanford professor Fei-Fei Li mass-harvested more than 14 million images from the internet and used Amazon Mechanical Turk workers to label them, importing cruel and racist categories from WordNet's lexical database. She documents the erosion of consent in data collection through examples including Duke University's dataset harvested from more than 2,000 students without their knowledge, later found to be used by the Chinese government for surveillance of ethnic minorities. Crawford argues that terms like "data mining" and "data is the new oil" reframed data from something personal to something inert and extractable.

The fourth chapter examines classification, beginning at the Penn Museum in Philadelphia, where Crawford visits almost 500 human skulls collected by nineteenth-century naturalist Samuel Morton, who measured them to rank human races. She argues that Morton's legacy foreshadows problems in AI: Correlating physical characteristics with intelligence acts as a technical alibi for oppression. She catalogs discriminatory AI systems, including Amazon's 2014 experiment automating hiring, in which a model trained on 10 years of résumés actively downgraded applications from women's college graduates. She examines IBM's Diversity in Faces dataset, which measured craniofacial distances using binary gender labels and claimed that heritage, race, and ethnicity are "reflected in our faces." ImageNet's "Person" category contained 2,832 subcategories including offensive labels that remained for 10 years until ImageNet Roulette, a public project that exposed the dataset's harmful labels, brought attention to the problem.

The fifth chapter traces the history of affect recognition. Crawford describes psychologist Paul Ekman's 1967 arrival among the Fore, an Indigenous people of Papua New Guinea, where he attempted to prove that facial expressions reveal universal emotions. His research was funded by the Advanced Research Projects Agency (ARPA), a research arm of the Department of Defense. Crawford catalogs the current affect recognition industry, including the hiring platform HireVue's use of facial cues to assess job candidates and the startup Affectiva's database of over 10 million expressions from 87 countries. She details scientific critiques culminating in neuroscientist Lisa Feldman Barrett's conclusion that it is "premature to use this technology to reach conclusions about what people feel on the basis of their facial movements." Despite this evidence, the industry continues expanding because powerful institutional and corporate investments depend on its validity.

The sixth chapter examines AI as a tool of state power, beginning with classified programs from the Snowden archive. Crawford traces AI's military origins, noting that the Defense Advanced Research Projects Agency (DARPA) was the primary patron for the first 20 years of AI research. She details Project Maven, which contracted Google to use TensorFlow, its machine-learning software framework, to analyze drone footage; more than 3,100 employees protested, and Google withdrew. Microsoft subsequently won the 10-billion-dollar Joint Enterprise Defense Infrastructure (JEDI) contract, a Pentagon cloud-computing deal. Crawford examines Palantir, the surveillance company cofounded by PayPal billionaire Peter Thiel. Its FALCON app enabled U.S. Immigration and Customs Enforcement (ICE) agents to pull data from multiple databases for immigration raids. Sociologist Sarah Brayne found that Palantir's platforms subjected predominantly poor, Black, and Latinx neighborhoods to heightened surveillance. Crawford describes IBM's experimental "terrorist credit score" system, which harvested Twitter data and lists of drowning victims during the 2015 Syrian refugee crisis to generate hypothetical threat scores for refugees without their knowledge. She details how AI-driven austerity programs in Michigan falsely accused thousands of residents of fraud.

In the conclusion, Crawford critiques what she and historian Alex Campolo call "enchanted determinism," the dual tendency to view AI as either a utopian solution or a dystopian overlord, arguing that both perspectives ignore systemic forces of inequality and exploitation. She contends that AI ethics principles are necessary but not sufficient, noting that 128 frameworks existed in Europe alone by 2019, yet these are rarely enforceable. Crawford calls for a politics of refusal, opposing technological inevitability and uniting movements for climate justice, labor rights, racial justice, and data protection. A brief coda examines tech billionaires' space ambitions, from Amazon founder Jeff Bezos's Blue Origin to Elon Musk's SpaceX, as extensions of the extractive logics analyzed throughout the book.