Redistributed

Book: Kate Crawford. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Link: Yale University Press

Why This Book Matters

The AI industry sells a story about itself: artificial intelligence is digital, immaterial, ethereal. It lives in the cloud. It runs on math. It emerges from data. This framing makes AI seem weightless, inevitable, almost natural — as if intelligence simply appeared when enough compute accumulated.

Kate Crawford wrote this book to demolish that story.

Atlas of AI traces what lies behind the screen: lithium mines in Nevada, cobalt extraction in the Congo, rare earth processing in Inner Mongolia, data labeling operations in India, warehouse workers monitored by algorithms in Kentucky. Crawford shows that AI is not virtual but viciously material. Every model runs on hardware. Every hardware component was extracted from the earth. Every training dataset was labeled by human hands. Every deployment reshapes human labor.

The book's power lies in this insistence on materiality. By following AI's supply chains from ground to cloud, Crawford reveals what the industry hides: the labor, the resources, the environmental costs, and the populations whose lives are shaped by systems they never consented to. The "atlas" metaphor is precise — this is cartography, mapping terrain that is deliberately kept invisible.

For redistribution, this is foundational. Crawford shows that AI doesn't create value from nothing; it extracts value from specific places and people. The wealth flows from periphery to center: from miners to refineries to factories to data centers to companies in San Francisco and Seattle. Understanding this geography is prerequisite to changing it.

The Argument

Crawford structures the book as six expeditions, each mapping a different dimension of AI's hidden infrastructure. Together they form a comprehensive challenge to how we understand artificial intelligence.

Earth: The Extractive Origins

The book opens at the Clayton Valley lithium mine in Nevada. This is where AI begins — not with code but with geology. Lithium-ion batteries power the devices that run AI. The batteries require lithium, cobalt, nickel, manganese. These materials come from specific places, extracted through specific processes.

Crawford documents what extraction means: Open pits. Toxic processing. Water contamination. Displaced communities. The environmental devastation at rare earth processing sites in Baotou, Inner Mongolia. The cobalt mines of the Democratic Republic of Congo, where workers including children dig without protective equipment.

The redistribution lens: The environmental and health costs of AI infrastructure are borne by specific communities, mostly in the Global South. The benefits accrue to users and companies mostly in the Global North. This is not an externality to be managed later; it's the foundation on which AI is built.

Labor: The Hidden Human Work

The second chapter traces human labor at every stage of AI development. Data labeling. Content moderation. Warehouse picking guided by algorithmic management. Gig workers handling tasks that automation can't.

Crawford visits an Amazon fulfillment center. Workers' movements are tracked continuously. Algorithms set pace targets. Bathroom breaks threaten productivity metrics. This isn't AI replacing human labor — it's AI intensifying human labor, extracting more work per hour while transferring risk and precarity to workers. The dynamics Crawford documents here have only accelerated, as The Manager Is an Algorithm explores in the context of contemporary logistics and gig work.

The chapter connects to Gray and Suri's Ghost Work, published two years earlier. Crawford extends their analysis: the invisible labor isn't just data labeling but the entire supply chain. Foxconn workers assembling devices. Logistics workers moving packages. Content moderators viewing traumatic imagery. Each link in the chain involves humans whose contributions disappear from the story of "artificial" intelligence.

The redistribution lens: Labor is redistributed from visible, protected, relatively well-compensated positions to invisible, precarious, poorly paid ones. The fiction of automation — that AI does the work — obscures this transfer. Workers remain essential; their conditions worsen.

Data: The Extraction of Human Lives

The third chapter examines where training data comes from. The answer: everywhere, often without consent.

Crawford traces the history of AI training datasets. ImageNet, the foundational image dataset, was built by scraping the web and using underpaid Mechanical Turk workers to label millions of images. The categories came from WordNet, a 1980s linguistic project with its own embedded assumptions. People's photos — their faces, their children, their private moments — became training data without their knowledge.

But data extraction goes beyond images. Every click, every search, every purchase generates data that feeds AI systems. Social media posts become training corpora. Scraped websites become language model inputs. The internet's content, created by billions of people, becomes raw material for systems those people neither control nor benefit from. The question of Who Owns Training Data remains legally and ethically unresolved.

Crawford extends the analysis to more explicit extraction. Surveillance systems collect data from populations who have no choice. Facial recognition databases built from mugshots. Emotion detection systems trained on images of children. Each dataset embeds assumptions about who can be watched, who can be used.

The redistribution lens: People generate data through their daily lives. That data creates value. The value flows to companies, not to the people who generated it. This is extraction — taking something from many, concentrating it in few.

Classification: Imposing Categories on the World

The fourth chapter may be the book's most philosophically challenging. Crawford argues that AI systems don't just process categories — they impose them.

Every classification system embeds choices. What are the categories? Who defined them? What does membership mean? These choices aren't neutral; they reflect the perspectives of those who made them.

Crawford examines historical classification systems: phrenology, criminal anthropology, racial typologies. These pseudosciences claimed to read character from physical features. They were used to justify discrimination, incarceration, exclusion. They were wrong — not just empirically but conceptually. The categories didn't exist in nature; they were imposed.

Modern AI classification systems repeat this pattern. Facial recognition that sorts people by race. Emotion detection that claims to read feeling from expression. Hiring algorithms that classify candidates as good or bad fits. Each system takes a human-defined category and treats it as objective reality.

The problem isn't just accuracy; it's ontology. Classification systems don't discover pre-existing categories — they create categories and then apply them to people. When the categories are consequential (who gets hired, who gets surveilled, who gets benefits), the power to define them is the power to shape lives.

The redistribution lens: The power to classify is distributed unequally. Companies and governments define the categories. Populations are sorted into them. Those classified rarely have voice in how categories are defined or applied. This is a redistribution of epistemic power — who gets to say what counts as what.

Affect: The Dubious Science of Emotion Recognition

The fifth chapter focuses on a specific domain: affect recognition, the claim that AI can detect emotions from faces.

Crawford traces affect recognition's intellectual lineage to Paul Ekman's theory of basic emotions — the claim that six fundamental emotions (happiness, sadness, fear, anger, surprise, disgust) are universal and expressed through consistent facial configurations. Ekman's work was influential. It shaped everything from airport security to job interviews to classroom monitoring.

The problem: the science doesn't hold up. A 2019 meta-analysis by Lisa Feldman Barrett and colleagues found that facial expressions don't reliably map to emotional states. Context matters. Cultural norms matter. Individual variation matters. The "basic emotions" framework was an oversimplification that decades of research have undermined.

Yet affect recognition systems continue to be built and deployed. Amazon's Rekognition. Microsoft's Emotion API. HireVue's video interview analysis. These systems claim to read candidates' emotions, assess students' engagement, detect suspicious behavior. They're built on questionable science, but they're used to make consequential decisions.

Crawford's analysis is devastating: affect recognition sells certainty it cannot deliver. The systems claim to know what you're feeling. They're probably wrong. But they're deployed in contexts where being wrong harms people — job applicants rejected, students flagged, travelers detained.

The redistribution lens: The costs of affect recognition's errors fall on those subjected to it. The benefits accrue to those deploying it (efficiency, the appearance of objectivity). This redistributes risk from institutions to individuals, from those with power to those without.

State: AI in the Service of Control

The final chapter examines how states use AI for surveillance, border control, and policing.

Crawford documents the integration of AI into state power: facial recognition for law enforcement, predictive policing algorithms, automated border control, welfare fraud detection, military targeting systems. In each case, AI amplifies the state's ability to watch, classify, and act on populations.

The chapter examines specific deployments. Palantir's work with Immigration and Customs Enforcement. The use of AI in drone targeting. China's surveillance infrastructure in Xinjiang. Predictive policing systems deployed in American cities. Each example shows AI extending state power over specific populations.

Crawford connects this to historical patterns. States have always sought to make populations legible — countable, categorizable, controllable. AI accelerates this project. It processes more data, operates at scale, makes decisions faster. What was once impossible (continuous surveillance of a city) becomes routine.

But the distribution isn't uniform. Some populations are watched more than others. Poor neighborhoods. Immigrant communities. Racialized groups. AI surveillance concentrates on those already marginalized, compounding existing inequalities.

The redistribution lens: AI redistributes power toward the state and away from individuals, particularly marginalized individuals. Surveillance is not evenly distributed; it concentrates on those with less ability to resist. The state gains capability; affected populations lose privacy, autonomy, and recourse.

The Map as Methodology

Crawford's method — following supply chains, visiting sites, drawing connections — is itself an argument. AI is usually presented abstracted from its material conditions. By mapping those conditions, Crawford refuses the abstraction.

The atlas metaphor is apt. An atlas reveals terrain: distances, elevations, borders, routes. Crawford's atlas reveals AI's terrain: where materials come from, where labor happens, where data flows, where power concentrates. The map makes visible what the industry prefers invisible.

This is a political choice. If AI is ethereal, questions of extraction don't arise. If AI is material, we can ask: Who mines the lithium? Who labels the data? Who bears the environmental costs? The map enables questions that the industry's preferred framing suppresses.

What the Book Gets Right

Materiality as corrective. The book's core move — insisting on AI's physical and human infrastructure — is essential. Too much AI discourse accepts the industry's framing of intelligence emerging from math. Crawford shows the mines, the factories, the warehouses, the workers.

Supply chain analysis. Following the supply chain is powerful method. It reveals dependencies, costs, and power relations that don't appear in company presentations. Crawford learned from scholars like Anna Tsing (The Mushroom at the End of the World) and applied the approach to AI.

Historical continuity. Crawford doesn't treat AI as unprecedented. She connects affect recognition to phrenology, surveillance to colonial administration, extraction to prior resource extraction. This continuity reveals patterns: new technologies serving old power dynamics.

Interdisciplinary synthesis. The book draws on environmental science, labor history, philosophy of science, surveillance studies, and more. This breadth is necessary. AI's impacts can't be understood from a single disciplinary perspective.

Accessible without simplifying. Crawford writes clearly for general audiences without dumbing down complex arguments. This is harder than it looks.

Critiques and Limitations

Solutions remain underdeveloped. Crawford excels at diagnosis. The analysis of what AI is and how it works is rigorous and original. But what to do about it remains less developed. The conclusion gestures toward regulation, labor protections, and challenging concentrated power, but without the specificity of the critique. Readers looking for constructive paths forward might pair this book with the policy frameworks in Power and Progress or the governance models explored in Democratic AI Governance.

Geographic limitations. The book focuses primarily on US and Chinese AI development. Important work, but AI development is increasingly global. India, Israel, the UAE, and others are significant players whose dynamics differ from the US-China axis.

Technological evolution. Published in 2021, the book predates ChatGPT and the generative AI explosion. The core argument about materiality holds — LLMs still require compute, which requires hardware, which requires extraction — but the specific applications have shifted. RLHF labor, for instance, receives less attention than it would in a 2024 book.

Corporate differentiation. Crawford sometimes treats "the AI industry" as monolithic. In practice, companies differ in practices, governance, and openness to reform. Some distinctions might illuminate what makes change possible.

The atlas metaphor's limits. Maps can suggest static terrain. AI's geography shifts rapidly. New mining sites, new labor arrangements, new data sources, new applications. The atlas approach is powerful but may understate dynamism.

What's Changed Since 2021

The book was published as the AI industry was about to transform. Several developments warrant update:

Generative AI's scale. The compute required for frontier models has increased dramatically. GPT-4 training is estimated at 100x GPT-3's cost. The environmental and material implications have grown proportionally.

RLHF labor. Reinforcement learning from human feedback has created new labor categories. Workers at Scale AI, Sama, and other firms rate outputs, write prompts, and shape model behavior. The conditions echo what Crawford documented for earlier AI labor — low pay, precarious classification, hidden from the product's public face. For more on this evolving labor category, see RLHF Workforce and Data Labeling Global South.

Lithium and supply chain pressure. Demand for lithium and other battery materials has intensified. New mines have been proposed and contested. Crawford's analysis of extraction has become more urgent, not less.

Regulatory movement. The EU AI Act, executive orders on AI safety, and various national initiatives have begun to address some concerns Crawford raises. Classification systems face more scrutiny. High-risk applications require more documentation. Whether these regulations are adequate remains contested.

AI critique has gone mainstream. When Crawford wrote, mapping AI's hidden costs was relatively novel. Now it's a recognized genre. This success creates its own challenges: how to move from critique to change.

Engaging with the Broader Literature

Atlas of AI belongs to a rich body of work on technology and power:

Langdon Winner's "Do Artifacts Have Politics?" asked whether technologies embed political choices. Crawford's answer is yes — not just in design choices but in supply chains, labor relations, and data practices.

Gray and Suri's Ghost Work documented hidden labor in AI specifically. Crawford expands the analysis to the entire supply chain — not just data labelers but miners, assemblers, and warehouse workers.

Eubanks's Automating Inequality showed how AI harms the poor. Crawford's chapter on state uses extends this: AI doesn't just automate existing inequality; it enables new forms of surveillance and control.

Benjamin's Race After Technology analyzed how technology encodes racial hierarchy. Crawford's classification chapter draws on this tradition, showing how AI categories impose racial and other hierarchies.

Srnicek's Platform Capitalism analyzed how platforms extract value. Crawford applies similar logic: AI extracts value from materials, labor, and data, concentrating it in few companies.

Implications for Redistribution

Crawford's atlas maps redistribution in multiple dimensions:

Material extraction. Lithium, cobalt, rare earths flow from extraction sites to manufacturers to data centers. Value concentrates in companies; environmental and health costs remain at extraction sites. This is geographic redistribution from periphery to center.

Labor extraction. Human work at every stage — mining, manufacturing, labeling, moderating — creates value that flows to AI companies. Workers' contributions are hidden; companies claim credit for "artificial" intelligence. This is redistribution from labor to capital.

Data extraction. People's digital lives become training data. The value created flows to companies, not to data sources. This is redistribution from the many to the few.

Power concentration. AI systems amplify state and corporate power. Surveillance, classification, and automated decision-making extend control over populations. Power redistributes from individuals to institutions.

Risk redistribution. Environmental risks remain at extraction sites. Labor risks fall on precarious workers. Classification risks harm those classified. Benefits concentrate; costs disperse.

Crawford's map reveals that "artificial intelligence" is a redistribution mechanism. It takes from many — from the earth, from workers, from data generators, from classified populations — and concentrates value in few. Understanding this geography is the first step toward changing it.

How to Use This Book

When discussing AI's environmental impact. Crawford provides concrete supply chain analysis. Not "AI uses energy" but "this lithium was mined here, processed there, assembled here, powered by this."

When challenging claims about AI's virtual nature. The book documents materiality in detail. Use it against framings that treat AI as weightless code.

When analyzing classification systems. Crawford's argument that classification imposes categories rather than discovering them has wide application. Any AI that sorts people into categories raises the questions she raises.

When examining AI labor. The book maps labor across the supply chain. Use it to extend analysis beyond data labeling to manufacturing, logistics, and operations.

When thinking about AI governance. Crawford shows why governance must address material conditions, not just algorithms. Supply chain transparency, labor standards, and environmental regulation are AI governance.

Source

Crawford, Kate. Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence. Yale University Press, 2021. Yale University Press

General

Book Review: Atlas of AI