Book summary “Atlas of AI”: summary

Introduction

You interact with artificial intelligence daily, yet the underlying systems appear invisible, weightless, and objective. When you use digital services, you rely on technologies marketed as immaterial code operating independently in the cloud. This framing obscures the immense physical infrastructure, human exploitation, and political power required to make computation function. You miss how these systems extract resources, classify populations, and reinforce inequality.

Kate Crawford argues in Atlas of AI that artificial intelligence is neither artificial nor intelligent. Instead, it functions as an extractive industry built on planetary resources, human labor, and historical regimes of power. The book traces this technological apparatus from deep geological mines to automated state surveillance. You discover how computation depends on physical extraction, hidden human toil, unconsenting data harvesting, and biased taxonomies.

The book answers several vital questions about this technology:

• How do global supply chains fuel modern computing infrastructure?
• Who performs the unseen labor required to train machine learning models?
• What political and social biases are embedded into training datasets?
• How do automated classification systems reinforce historical prejudices?
• In what ways do state and corporate powers weaponize surveillance systems?

1. Earth

Artificial intelligence is a physical, extractive technology dependent on finite geological resources and fossil fuels rather than an immaterial cloud.

Planetary Extraction

Computation requires massive quantities of nonrenewable materials that took billions of years to form inside the earth. The tech industry subjects these geological elements to rapid extraction, processing, and transportation, ultimately resulting in short-lived consumer devices that end up in e-waste dumps.

The Consumer Technology Association notes that the average smartphone lifespan is 4.7 years. A smartphone battery typically contains 0.3 ounces of lithium. A Tesla Model S battery pack needs 138 pounds of lithium.

Tesla relies heavily on the Silver Peak lithium mine and consumes over 28,000 tons of lithium hydroxide annually, accounting for half of the planet's total consumption. Albemarle Corporation acquired Rockwood Holdings for 6.2 billion dollars in 2014 to secure these supplies.

Tech hardware manufacturers maintain plausible deniability regarding human suffering and environmental destruction by utilizing complex supply chains with opaque sourcing. Intel operates a supply chain with over 16,000 suppliers across more than 100 countries and required over four years to gain basic insight into its sources.

The Myth of Clean Tech

The production of rare earth elements essential for modern electronics generates extreme ratios of toxic waste relative to usable materials. Extracting elements like neodymium and dysprosium requires heavy acid washing, which yields massive reservoirs of radioactive residue and poisonous wastewater stored in artificial lakes.

In Baotou, Inner Mongolia, an artificial lake spanning over 5.5 miles in diameter holds more than 180 million tons of toxic black mud and waste powder from ore processing. Refining rare earth elements requires 75,000 liters of acidic water per ton, and produces 1 ton of radioactive residue per ton of refined materials. Mining in Jiangxi yields clay where only 0.2 percent contains valuable rare earth elements, leaving 99.8 percent of removed earth discarded as tailings.

Historical communication infrastructures followed identical patterns of ecological collapse and colonial exploitation driven by material demands. Laying the first transatlantic cable in 1857 required approximately 250 tons of gutta-percha, which meant stripping roughly 900,000 tree trunks per ton of processed material.

The massive computational power required to train and run modern artificial intelligence models generates a massive and rapidly growing carbon footprint. Researchers at the University of Massachusetts Amherst found that training a single natural language processing model produced over 660,000 pounds of carbon dioxide emissions.

Global logistical networks, standardized cargo transport, and massive data centers impose severe hidden environmental costs that contradict the image of a clean digital cloud. Container shipping relies on cheap, low-grade bunker fuels that emit massive amounts of sulfur and carbon dioxide, while large data centers consume millions of gallons of local water resources for cooling. The National Security Agency data center in Bluffdale, Utah, is a 1.2 million square foot facility designed to consume an estimated 1.7 million gallons of water per day, secured through a discounted water deal with the local city.

2. Labor

Large-scale computation relies entirely on the hidden toil of human workers across global supply chains, warehouses, and crowdsourcing platforms.

Corporate narratives promote the illusion of fully autonomous artificial intelligence. In reality, these systems rely on massive amounts of unseen human toil. This includes supply-chain mining, software microtasks, and physical labor in distribution warehouses.

Amazon fulfillment centers use Kiva robots for heavy lifting. Human workers are still required as connective tissue to pick and pack oddly shaped objects.

Many systems marketed as artificial intelligence are actually illusions. They are powered by hidden armies of low-paid human workers performing repetitive tasks behind the scenes. Astra Taylor terms this deceptive practice fauxtomation, while others call it Potemkin AI or artificial artificial intelligence.

The personal assistant startup x.ai claimed an AI named Amy scheduled meetings. Human contract workers actually checked and rewrote emails manually during long shifts.

The historical hoax of a mechanical chess player operated by a hidden human master persists today. In 1770, Wolfgang von Kempelen constructed the mechanical chess player. It used a hidden human chess master inside a cabinet.

Amazon revived this exact concept about 250 years later with its Mechanical Turk platform. This system connects businesses to an anonymous mass of workers performing microtasks. Amazon created the platform after its own algorithms failed to adequately detect duplicate product pages on its retail site.

Charles Babbage conceptualized the industrial factory as a calculating machine. He viewed human workers merely as an unreliable source of error rather than a source of value. Building on Adam Smith's division of labor, Babbage developed the idea for the Difference Engine in the 1820s. Later in the 1830s, he designed the Analytical Engine.

Ghost Work and Fauxtomation

The origins of modern surveillance society lie in early industrial manufacturing facilities designed to monitor workers. In the 1780s, Samuel Bentham developed the inspection house in Russia. He used this design to keep workers within sight of supervisors from a central platform.

Samuel Bentham used the inspection house in Russia to monitor Russian peasants loaned by Prince Potemkin and to keep an eye on British master shipbuilders. This factory design later inspired his brother Jeremy Bentham to conceptualize the panoptic prison.

The Chicago meat-packing industry operationalized the disassembly line in the 1870s. Trains brought livestock to stockyards where carcasses moved along mechanized overhead trolleys. This complete division of labor turned workers into interchangeable, low-skill commodities performing micro-subdivided tasks.

Upton Sinclair wrote the novel The Jungle to highlight working-class poverty in Chicago meat-packing plants. Public outcry focused instead on food safety and resulted in the Meat Inspection Act of 1906.

Frederick Winslow Taylor used stopwatches to micromanage the microscopic physical movements of workers. His Principles of Scientific Management quantified human bodily movements to establish optimal layouts for tools and processes. This system maximized output at minimal cost while treating workers as organic machines subservient to clock time.

Large-scale technological infrastructures create proprietary global time protocols to synchronize data centers across the planet. Google's Spanner database uses a distributed time protocol called TrueTime, relying on GPS and atomic clocks. When individual server clocks drift, Spanner slows down time to maintain a unified sequence of events.

Setting the Rate

At Amazon fulfillment centers, algorithmic productivity targets known as the rate drive workers to physical exhaustion. This forms an inflexible core business model that management refuses to negotiate. Workers are continuously monitored and tracked against strict picking and packing rates.

Failing to meet the rate multiple times results in termination. Workers must skip bathroom breaks and suffer chronic injuries to keep pace.

An Amazon fulfillment center shift lasts 10 hours. Associates receive a 30-minute unpaid meal break and a maximum of 15 minutes of off-task time allowed per shift. Kiva robot towers carry weights up to 3,000 pounds.

Warehouse workers in Minnesota and California staged walkouts and protests. They protested unsustainable work quotas, bereavement leave restrictions, and carbon footprints.

3. Data

The massive acquisition of training data operates through unauthorized scraping, unconsenting capture of public commons, and extractive colonial logics.

Supervised machine learning relies on training datasets as ground truth. Because models make inductive inferences rather than deductive ones, any gaps or lack of diversity in the training set become absolute system rules.

Teaching a neural network to distinguish apples and oranges via supervised learning where all sample apples are red leads the model to deduce that all apples are red.

Capturing the Face

The FERET program established early standards for facial recognition training data. Sponsored by the Department of Defense CounterDrug Technology Development Program Office and led by the U.S. Army Research Laboratory, FERET collected high-resolution headshots of over a thousand people with full consent, creating a standardized measuring tool before the internet era of mass scraping.

The FERET program photographed subjects at George Mason University between 1993 and 1994 using 35 mm cameras in formal portraiture style with full ethics review board approval.

The dataset included 1,000+ people included in the FERET training set, 14,126 total images in the FERET dataset, and the 1996 year of the FERET research paper admitting distribution questions.

Mug shots from state archives are repurposed as training data for commercial and academic facial recognition AI. NIST maintains collections like the Multiple Encounter Dataset containing thousands of arrest photos, including deceased people and individuals showing signs of violence.

Kate Crawford discovered mug shot databases like NIST Special Dataset 32 in 2017, noting images of people with visible wounds, bruising, and distress used purely as technical substrates for Facial Recognition Verification Testing. Special Dataset 32 contains thousands of photographs of people with multiple arrests.

The Capture of the Commons

AI research largely operates without ethical review boards because computer science has historically been treated as abstract applied mathematics rather than human subjects research. Institutional review boards focus primarily on biomedical and psychological experimentation, leaving machine learning researchers exempt from scrutiny.

Harvard computer scientist Hau Chan presented a gang-crime prediction neural network trained on skewed Los Angeles Police Department and CalGang data, stating he was just a researcher who did not know how the tool would be used.

The AI industry engages in privatization by stealth, enclosing public goods and commons to extract private profits and insights without returning value to the public.

Britain's Royal Free National Health Service Foundation Trust shared patient data records of 1.6 million people with Google subsidiary DeepMind, violating data protection laws by failing to sufficiently inform patients.

4. Classification

Automated classification systems and training set taxonomies encode historical prejudice and flatten human complexity into rigid, discriminatory proxies.

As Bowker and Star observe, classifications tend to disappear into habit and infrastructure. In artificial intelligence, the categories casually chosen to structure technical systems play an active role in reinforcing social hierarchies and power dynamics.

Samuel Morton amassed over a thousand skulls, measuring them by filling cranial cavities with lead shot and publishing tables claiming white people had the largest skulls while Black people were at the bottom.

Historical efforts to rank human races through craniometry reveal that classification systems are deeply political exercises shaped by prior prejudice rather than objective science. Morton collected almost five hundred skulls collected in the early decades of the 1800s and more than a thousand skulls amassed by the end of Morton's life in 1851, using five races of the world in Morton's taxonomy.

Systems of Circular Logic

Machine learning models trained on historical data create a statistical ouroboros that reinforces past inequalities under the illusion of technical objectivity. When technical systems extract data from socially unequal environments and use it for pattern recognition, the resulting automated decisions perpetuate and intensify historical biases while presenting themselves as neutral outputs.

Amazon created an experimental automated hiring tool trained on ten years' worth of résumés used to build Amazon's hiring model containing fifty thousand terms that appeared in those résumés, which quickly learned to penalize female candidates because past engineering hires were predominantly men.

Large-scale training sets like ImageNet organize the world through labyrinthine hierarchies that flatten complex social and cultural phenomena into discrete, ideologically loaded categories. ImageNet imported its semantic structure from WordNet, first developed at Princeton University in 1985, organizing millions of images into 21,841 categories originally in the ImageNet hierarchy and miscellaneous top-level categories. Its person subcategories incorporated extensive moral judgments, racial identities, professions, and slurs, treating humans as objects with 2,832 subcategories under the top-level category Person.

ImageNet contained thousands of person subcategories including offensive terms, racial labels, and nonvisual relational concepts like debtor or color-blind person, which remained hidden until projects like ImageNet Roulette exposed them, leading the ImageNet team to remove 1,593 of 2,832 Person categories removed as unsafe by the ImageNet team, representing roughly 56 percent, alongside 600,040 images removed alongside the unsafe categories.

Constructing Race and Gender

Machine learning systems actively construct race and gender by treating them as fixed, natural, and automatically detectable biological categories.

The UTKFace dataset contains more twenty thousand images of faces in the UTKFace dataset annotated with a forced gender binary and five classes used to categorize race in UTKFace to automate age, ranging across zero to 116 years range for age estimation in UTKFace, and face detection tasks.

IBM created the Diversity in Faces dataset drawing from one hundred million images taken from Flickr as a preexisting dataset to measure craniofacial distances, using a small sample of one million photos used as a small sample for craniofacial measurements and crowdworkers to annotate age and gender while removing individuals who did not fit a binary gender model. Projects aimed at improving facial recognition accuracy frequently reduce diversity to mere cranial variations while enforcing rigid binary categories and ignoring transgender or nonbinary people.

5. Affect

Commercial affect recognition industries build on flawed psychological premises and military-funded research to police human expressions and interior states.

In 1967, psychologist Paul Ekman traveled to Papua New Guinea with flashcards of facial expressions to test his hypothesis on the isolated Fore people of Okapa. Ekman sought to prove universality by showing photographs of facial expressions to isolated indigenous groups. His work was heavily funded by the military during the Cold War and assumed that interior feelings are involuntarily revealed on the face and can be categorized into distinct, universal expressions.

Silvan Tomkins published his magnum opus, Affect Imagery Consciousness, in 1962, which heavily influenced Paul Ekman's career. Tomkins argued that while biological drives address immediate needs like hunger, affects govern general motivation and behavior. Because affects are ambiguous, he proposed that they have a one-to-one relationship with facial expressions.

Military Roots of AI Affect Detection

The Advanced Research Projects Agency funded Ekman's research largely because its division head needed to distribute money quickly to avoid political scrutiny regarding covert social science research in Chile. ARPA head Lee Hough wrote the research proposal that funded Ekman's cross-cultural studies to quickly disburse funds and avoid suspicion from Senator Frank Church after covert operations in Chile. This injection amounted to one million dollars, equivalent to more than eight million dollars today, seeding an expanding seventeen-billion-dollar industry built on Ekman's presuppositions.

Duchenne worked at the Salpetriere asylum in Paris, using electrical shocks on vulnerable patients' faces to simulate expressions and capture them using collodion photography among the five thousand people housed there. Early researchers like Lavater used silhouettes and skull measurements to link appearance with character in physiognomy and phrenology.

The Politics of Faces

Companies like HireVue, Human, and Affectiva use machine learning to assess job candidates or measure consumer emotions, with Affectiva collecting an emotion database made up of over ten million people's expressions from eighty-seven countries. Tech startups and major tech giants have built platforms that analyze video interviews, microexpressions, and tone of voice to score candidates or screen travelers. The SPOT passenger screening program carried a nine-hundred-million-dollar price tag while producing no clear successes.

Because datasets like the Cohn-Kanade and Karolinska databases rely heavily on actors and subjects simulating extreme expressions or responding to media while watched, AI systems are trained to recognize artificial performances rather than ground-truth interior states. Researchers at a Swedish lab created the Karolinska Directed Emotional Faces database, featuring extreme posed expressions corresponding to Ekman's categories like incredible surprise and paralyzing fear.

In 2019, Lisa Feldman Barrett led a research team that conducted a comprehensive review of the scientific literature, concluding that facial expressions are not reliable fingerprints or diagnostic displays of emotion. Interdisciplinary critics, including anthropologist Margaret Mead, historian Ruth Leys, and psychologist Lisa Feldman Barrett, have argued that Ekman's theories are recursive, oversimplified, and ignore the complex cultural conditioning of emotional display.

A conservative blog claimed to use a virtual polygraph system powered by Amazon Rekognition and IBM Watson to analyze Congresswoman Ilhan Abdullahi Omar, falsely labeling her a security threat and a pathological liar. Affect recognition tools risk weaponizing flawed technologies against marginalized groups by misinterpreting speech affects and facial expressions as indicators of deception or threat. Unrepresentative training data builds in racial and gender biases that frequently misinterpret normal stress, cultural differences, or minority faces as displaying negative emotions like anger or contempt.

6. State

Military agencies and intelligence commands shaped the foundational decades of artificial intelligence research. You find that objectives like command, automation, and target detection remain embedded in computer science. The Office of Naval Research partly funded the first Summer Research Project on Artificial Intelligence at Dartmouth College in 1956.

State intelligence agencies, military apparatuses, and municipal law enforcement integrate commercial AI to expand total surveillance and automate austerity.

Classified Empires and Surveillance

Leaked documents from the NSA reveal a parallel, secretive AI sector driven by total data extraction. An NSA PowerPoint deck outlined TREASUREMAP, a program designed to build a near real-time, interactive map of the internet and track the location and owner of any connected device.

The Department of Defense launched Project Maven in 2017 to use machine learning for analyzing drone footage. Over 3,100 Google employees signed a letter of protest demanding the cancellation of the Project Maven contract and that Google should not be in the business of war.

Palantir was established in 2004 and secured one billion dollars in total U.S. government contracts after Trump took the presidency. Immigration and Customs Enforcement used Palantir's mobile app, falcon, to guide a raid of nearly a hundred 7-Eleven stores across the United States. Vigilant Solutions signed a contract giving ICE access to five billion records of license plates gathered by private businesses and 1.5 billion data points from local law enforcement agencies. Amazon negotiated video-sharing partnerships with more than six hundred police departments, offering credits for free Ring cameras in exchange for promoting the app.

Algorithmic Austerity

Military intelligence relies on metadata signatures rather than verified identities to authorize lethal drone strikes. An estimated 1,147 people were killed by drone strikes attempting to target 41 individuals. IBM built an experimental terrorist credit score for border guards by harvesting unstructured social media data and lists of drowned refugees to weed out ISIS fighters from asylum seekers.

Automated systems deployed within the welfare state function as punitive mechanisms of exclusion and budget cutting. Michigan's MiDAS system inaccurately identified over forty thousand residents of suspected unemployment insurance fraud, leading to wage garnishments and severe civil penalties. A fugitive felon policy led to 19,000 Michigan residents being incorrectly disqualified from food assistance via fugitive felon matching.

Tech billionaires are funding a private space race aimed at off-world extraction. Jeff Bezos owned 300,000 acres of land in West Texas and founded Blue Origin to build reusable rockets and space colonies. Commercial space companies were granted regulatory exemption under the Commercial Space Launch Competitiveness Act up to the year 2023.

Conclusion

Planetary extraction, hidden human labor, unconsenting data capture, discriminatory classification, affect policing, and state surveillance connect into a unified informatics of domination. You must recognize that artificial intelligence is an extractive political industry rather than a neutral technical tool. You can confront these systems by organizing across disciplines and demanding structural accountability from technology companies and states.

1. Refuse the myth of immaterial technology and trace every digital system back to its geological and human origins.
2. Support labor organizing and legal protections for data annotators, content moderators, and gig workers across global supply chains.
3. Challenge unconsenting data collection and demand robust rights to privacy, bodily autonomy, and data refusal.
4. Expose the historical and political assumptions embedded in classification taxonomies, affect recognition models, and algorithmic governance.
5. Build cross-border coalitions that link environmental justice, labor rights, and anti-surveillance movements into a connected politics of refusal.

10 Key Ideas

1. Artificial intelligence operates as an extractive industry dependent on finite geological resources and fossil fuels rather than an immaterial cloud.
2. Large-scale computation relies entirely on the hidden toil of human workers across global supply chains, warehouses, and crowdsourcing platforms.
3. The massive acquisition of training data operates through unauthorized scraping, unconsenting capture of the public commons, and extractive colonial logics.
4. Automated classification systems and training set taxonomies encode historical prejudice and flatten human complexity into rigid, discriminatory proxies.
5. Commercial affect recognition industries build on flawed psychological premises and military-funded research to police human expressions and interior states.
6. State intelligence agencies, military apparatuses, and municipal law enforcement integrate commercial AI to expand total surveillance and automate austerity.
7. The tech industry subjects nonrenewable geological elements to rapid extraction and processing while maintaining plausible deniability regarding human suffering through opaque supply chains.
8. Amazon fulfillment centers use algorithmic productivity targets known as the rate to drive workers to physical exhaustion while continuously monitoring their movements.
9. Machine learning models trained on historical data create a statistical ouroboros that reinforces past inequalities under the illusion of technical objectivity.
10. Automated systems deployed within the welfare state function as punitive mechanisms of exclusion and budget cutting, as seen when Michigan's MiDAS system inaccurately identified over forty thousand residents of suspected fraud.