1장
The Silicon Valley's Dirty Secret: AI's Hidden Material Reality
When you think of artificial intelligence, what comes to mind? Perhaps sleek robots, voice assistants that respond to your every command, or algorithms that can beat world champions at complex games. But behind this glossy veneer lies a far more complicated reality. Kate Crawford's "Atlas of AI" has become a landmark text for understanding the true nature of artificial intelligence, stripping away the marketing hype to reveal the material foundations of what we call AI. The book has been praised by tech ethicists and industry insiders alike for its unflinching examination of AI's hidden costs. Even Elon Musk, despite his own AI ventures, acknowledged its importance, tweeting that it offers "uncomfortable but necessary insights." Since its 2021 publication, it has become required reading in tech ethics courses at universities worldwide, offering a counternarrative to Silicon Valley's relentless techno-optimism. Crawford takes us on a journey from lithium mines in Nevada to crowdworker platforms in India, revealing that AI is neither artificial nor intelligent-but rather a massive industrial system built on extraction, exploitation, and classification that shapes our world in profound ways.
2장
Earth: The Hidden Material Foundations of "The Cloud"
The journey begins with an aerial view of Silicon Valley's gleaming tech campuses before shifting to Nevada's remote Clayton Valley, where massive green evaporation ponds extract lithium-the "gray gold" powering our devices. This geographical contrast reveals a fundamental truth: the technological marvels of Silicon Valley depend on extractive processes happening far from public view.
This connection isn't new. San Francisco was built from gold and silver extraction from lands taken from Mexico, with mining operations that devastated the Central Valley while creating enormous wealth. Today's technology industry follows the same pattern, externalizing environmental costs while concentrating profits. The world's five largest companies now headquartered in a city experiencing extreme inequality, where billionaires and homeless encampments exist in close proximity.
At Silver Peak, Nevada, the Albemarle Corporation pumps valuable brine from an underground lithium lake into open ponds to evaporate. This operation, acquired for $6.2 billion in 2014, represents the only operating lithium mine in the United States. Tesla alone consumes approximately half the planet's total lithium hydroxide production annually, making control of these resources strategically vital.
But Nevada's lithium is just one node in a global extraction network. From Bolivia's salt flats to Congo's coltan mines and Mongolia's rare earth processing facilities, AI's material foundation spans the globe. The U.S. Geological Survey identified 23 "high supply risk" minerals essential to manufacturing, including rare earth elements that make our devices smaller and lighter. Despite legislation like the Dodd-Frank Act regulating "conflict minerals," tech giants struggle to verify their supply chains. Companies like Intel, with 16,000+ suppliers across 100+ countries, took years to develop even basic supply chain visibility.
In Baotou, Inner Mongolia, a toxic black lake stretching over five miles contains 180 million tons of waste from rare earth processing. China supplies 95% of the world's rare earth minerals not due to geological advantage but willingness to absorb environmental damage. The extraction process dissolves minerals in sulfuric and nitric acid, creating vast reservoirs of poisonous waste. For elements like dysprosium and terbium, only 0.2% of mined clay contains valuable materials-the rest becomes waste dumped into hills and streams.
Despite its ethereal "cloud" metaphors, AI infrastructure is profoundly material and environmentally damaging. The tech sector projects an image of sustainability while concealing enormous energy consumption. Data centers already match aviation's carbon footprint and are growing faster, potentially contributing 14% of global emissions by 2040. A single natural language processing model produces 660,000 pounds of carbon dioxide, equivalent to five cars' lifetime emissions. Major tech companies claim carbon neutrality through offset credits while simultaneously licensing AI to fossil fuel companies.
Water consumption represents another hidden cost. The NSA's massive Utah data center requires 1.7 million gallons of water daily, yet initially refused to disclose usage data, claiming national security concerns. Like mining operations, most data centers are deliberately placed far from population centers, reinforcing the illusion that cloud computing is immaterial.
AI functions as what Lewis Mumford called a "megamachine"-an enormous system comprised of individual human actors working in secrecy and coordination. Like the Manhattan Project that employed 130,000 workers to develop atomic weapons, AI is metamorphic and opaque, relying on global manufacturing, transportation networks, data centers, undersea cables, personal devices, and scraped datasets. The scale and complexity of this material infrastructure is deliberately obscured by intellectual property laws and technical complexity.
3장
Labor: The Human Workers Behind "Automated" Systems
Amazon's vast fulfillment center in Robbinsville exemplifies how computational logics control workers' bodies and time. Inside the 1.2 million square foot warehouse, "associates" are monitored by time clocks, allowed only fifteen minutes off-task per ten-hour shift, and must pass through metal detectors when leaving. While orange Kiva robots glide efficiently carrying shelving units, human workers wearing knee braces and wrist supports struggle to meet demanding "picking rates," with vending machines stocked with painkillers throughout the facility.
Rather than debating whether robots will replace humans, we should examine how humans are increasingly treated like robots under AI systems that maximize labor extraction through surveillance and algorithmic assessment. This isn't a future concern but a long-established reality. Manufacturing assembly lines have spread to service industries, secretarial work has been automated since the 1980s, and even white-collar "knowledge workers" face increasing workplace surveillance and process automation.
The rhetoric of "human-AI collaboration" masks a significant power asymmetry-workers rarely have the option to opt out of algorithmic systems, instead being forced to re-skill and unquestioningly accept each new technical development. Behind AI's veneer of automation lies an invisible workforce of underpaid laborers. From miners extracting resources to crowdworkers tagging thousands of images for pennies per task, exploitative labor exists throughout the AI pipeline.
Content moderators suffer psychological trauma while reviewing violent content, yet their work remains essential for AI functionality. Companies like x.ai and Facebook have even deployed "Potemkin AI"-services marketed as automated that secretly rely on humans working exhausting shifts behind the scenes. Amazon's Mechanical Turk platform-named after an 18th century chess-playing hoax that hid a human operator-epitomizes this dynamic, with workers bidding against each other for microtasks while remaining invisible.
This approach has historical precedent. Charles Babbage, inventor of the first mechanical computer, envisioned factories as computational systems where specialized units performed coordinated tasks while rendering labor invisible. He saw value coming from manufacturing process design rather than workers, whom he viewed as potential sources of error. This vision materialized in Chicago's 1870s meat-packing industry, where the "disassembly line" reduced skilled butchery to simple, replaceable tasks. When Upton Sinclair's novel The Jungle exposed these conditions, powerful institutions addressed food safety concerns but ignored the underlying labor exploitation.
Modern workplaces have intensified time control to manage bodies. Ray Kroc's McDonald's system exemplifies this approach, removing "almost all conceptual work from execution of tasks" through detailed standardization of every movement. Today's employers use passive surveillance-access badges, fingerprint readers, timing devices, and workplace sensors tracking everything from body temperature to website browsing. Silicon Valley's predominantly young male workforce creates productivity tools that police other workplaces, establishing masculinized benchmarks that rely on others' underpaid care work.
Amazon workers are organizing across facilities to challenge "the rate"-the core productivity metric that Amazon executives refuse to negotiate. When Somali workers in Minnesota protested unsustainable working conditions, Amazon representatives declared "the rate is our business model. We cannot change that." Unlike local labor disputes where supervisors might make concessions, the rate is programmed into Amazon's computational infrastructure by distant executives. This has sparked cross-sector solidarity among diverse workers who recognize their shared struggle against algorithmic monitoring systems.
4장
Data: The Extraction and Exploitation of Digital Resources
The transformation of human images into mere technical resources exemplifies AI's extractive logic. NIST's mug shot databases-showing arrested individuals, many with visible injuries or distress-have become standardized testing data for facial recognition algorithms. These deeply personal images of vulnerable people have shifted from tools of identification to abstract training material stripped of context, consent, and humanity.
Computer vision is a profoundly complex endeavor, yet AI developers commonly scrape millions of images from the internet, classify them, and use these collections as "ground truth" for teaching machines to see. In supervised machine learning, human engineers supply labeled training data to algorithms: learners that train on examples, and classifiers that analyze new inputs. Training datasets define the epistemic boundaries of how AI can "see" the world, creating brittle foundations that inevitably simplify our infinitely complex reality into limited categories.
The hunger for data in computing has deep roots. In 1945, Vannevar Bush envisioned future machines with "enormous appetites" for data, processed by "roomful of girls" doing computation work. By the 1980s, AI research shifted from rules-based expert systems to probabilistic approaches requiring massive datasets. IBM's speech recognition breakthrough exemplifies this transformation, as researchers abandoned linguistic methods for statistical ones that treated speech as merely data. Their mantra became "There's no data like more data," leading them to desperately seek text from technical manuals, children's novels, and even antitrust lawsuit depositions.
The internet revolutionized AI training data collection, eliminating the need for controlled photo shoots as researchers began treating online content as a natural resource to harvest. With platforms like Facebook receiving 350 million daily photo uploads and Twitter processing 500 million tweets in 2019, tech companies gained valuable proprietary data troves. Academic researchers sought similar advantages, leading to projects like ImageNet. Started in 2006 by Professor Fei-Fei Li to "map out the entire world of objects," ImageNet extracted over 14 million images from the internet and organized them into 20,000+ categories.
When manual labeling by Princeton undergraduates proved too expensive, the team turned to Amazon Mechanical Turk, becoming its largest academic user. Workers sorted approximately fifty images per minute, including problematic categories that labeled people as "alcoholic," "crazy," and other offensive terms. This approach-mass extraction without consent combined with underpaid crowdworker labeling-became standard practice throughout the field.
The early 2000s marked a shift away from consent-driven data collection as researchers presumed internet content was theirs for the taking. Increasingly troubling extraction practices emerged: a University of Colorado professor secretly photographed 1,700+ students on campus walkways; Duke University's Army-funded project captured footage of 2,000+ students without knowledge; Stanford researchers commandeered a San Francisco cafe webcam to extract 12,000 images without consent.
The field has normalized mass data extraction through powerful myths and metaphors. Terms like "data mining" and "data is the new oil" rhetorically shift data away from something personal toward something inert and nonhuman-a resource to be extracted. This bloodless language disguises both data's material origins and consequences. Framing data as a "natural resource" waiting to be discovered employs colonial rhetoric that justifies extraction from "unrefined" sources.
Most university-based AI research operates without ethical review, stemming from the historical view that applied mathematics, statistics, and computer science didn't constitute human subjects research. Even as AI moved from laboratories into real-world applications affecting welfare benefits and criminal sentencing, datasets remained exempt from ethics review. This arm's-length relationship between researchers and subjects creates dangerous situations where harmful ideas propagate unchecked.
5장
Classification: The Politics of Categorizing the World
The chapter examines how AI bias reflects deeper classification problems. Discriminatory AI systems are now widely documented, from gender bias in credit algorithms to racism in criminal risk assessment software. The typical response follows a pattern: exposure by journalists, company promises to address the issue, and proprietary "fixes" with little public debate about fundamental problems. Amazon's hiring AI experiment serves as a vivid example-the system downgraded women candidates by learning from the company's male-dominated hiring history, essentially creating a diagnostic tool that revealed existing biases.
IBM's approach to diversity in facial recognition reveals problematic politics: they used binary gender categories and removed anyone outside this framework, reducing "fairness" to higher accuracy rates and "diversity" to mere facial variation. Their problematic claim that "aspects of our heritage and individual identity are reflected in our faces" contradicts decades of research showing race, gender, and identity are socially constructed, not biological categories.
Classification has historically been linked to power-from divine naming in theology to the Greek root of "category" suggesting both logical assertion and legal accusation. While "bias" began as a geometric term for diagonal lines, by the 1900s it gained a technical meaning in statistics as systematic differences between samples and populations. Technical solutions addressing statistical bias often fail to address deeper structural inequalities and power asymmetries that persist regardless of designers' intentions.
ImageNet's classification structure reveals its underlying worldview. Built on WordNet (a semantic database developed at Princeton in 1985 with U.S. Navy funding), ImageNet organizes nouns into "synsets" within a nested hierarchy. Its nine curious top-level categories-plant, geological formation, natural object, sport, artifact, fungus, person, animal, and miscellaneous-spawn into thousands of specific nested classes containing millions of images. Human classifications particularly reveal political assumptions: "human body" falls under "Natural Object," with subcategories limited to "male" and "female" bodies, while terms like "Hermaphrodite" appear under "Person -> Sensualist -> Bisexual."
The act of categorization reifies the existence of categories themselves. While classifying "apple" seems uncontroversial, nouns exist on a spectrum from concrete to abstract, descriptive to judgmental-distinctions erased in ImageNet's flattening taxonomy. For a decade, ImageNet contained 2,832 subcategories under "Person," with the most populated being "gal," "grandfather," "dad," and "CEO" (mostly male). The dataset also attempted to classify people by professions and relationships that aren't visually determinable, such as "Debtor," "Boss," "Acquaintance," and "Color-Blind Person"-nonsensical visual categories that nonetheless become reified through their connection to images.
Classification systems in AI falsely assume gender, race, and sexuality to be natural, fixed, and detectable biological categories. Datasets like UTKFace reduce people to binary gender (0 for male, 1 for female) and five racial categories (White, Black, Asian, Indian, Others)-a schema reminiscent of South Africa's apartheid classifications. These reductive taxonomies echo historical systems of oppression where IBM maintained databases for racial classification that governed people's lives and criminalized interracial relationships.
The fundamental question facing AI classification systems is who gets to choose how social categories are represented and on what basis. Justice in AI systems cannot be simply coded or computed-it requires moving beyond optimization metrics and statistical parity to understand where mathematical and engineering frameworks themselves cause problems. Classificatory infrastructures necessarily reduce complexity and remove context to make the world computable. The most harmful forms of human categorization throughout history didn't simply fade away under scientific scrutiny but required political organizing and sustained protest.
6장
Affect: The Myth of Reading Emotions from Faces
In a remote outpost in Papua New Guinea, American psychologist Paul Ekman arrived with flashcards to test his theory that humans exhibit universal emotions across cultures. This controversial hypothesis has grown into a multi-billion dollar affect recognition industry embedded in AI systems worldwide. Today, emotion detection technologies are deployed in hiring, security, education, and policing-all claiming to detect interior emotional states from facial expressions.
Despite their widespread adoption, these systems rest on shaky scientific ground. A comprehensive 2019 review found no reliable evidence that emotions can be accurately predicted from facial movements. Yet the allure of automated affect recognition remains powerful for militaries, corporations, and police forces seeking to distinguish friend from foe and extract hidden truths.
Technology companies have amassed vast repositories of human facial expressions through social media platforms, enabling attempts to extract "hidden" emotional states using machine learning. Companies like Human and HireVue use these systems to evaluate job candidates, while tech giants including Apple, Amazon, Microsoft, and IBM offer emotion detection tools claiming to identify "universal" emotions like anger, contempt, and happiness.
Ekman's research began with an influential encounter with psychologist Silvan Tomkins, whose work on affect theory proposed that emotions were innate evolutionary responses recognizable across cultures-an appealing simplification that made these theories easily replicable in AI systems. Critically, Tomkins acknowledged that while affects might be universal, their interpretation depends on social and cultural factors-creating "dialects" of facial language across societies.
Ekman's research direction was significantly shaped by funding from the Defense Department's Advanced Research Projects Agency (ARPA), which saw potential in understanding cross-cultural nonverbal communication. This began a pattern of defense, intelligence, and law enforcement agencies funding Ekman's career and the broader field of affect recognition.
The belief that internal states can be read from external appearances has deep historical roots in physiognomy. This practice, dating back to Aristotle's assumption that "body and soul are affected together," was also used for racial classification. French neurologist Duchenne revolutionized this field by using photography and electrical stimulation to study facial expressions. Working with asylum patients who couldn't refuse his experiments, Duchenne applied electrical shocks to isolate muscle movements, photographing the resulting expressions.
Ekman followed Duchenne's photographic approach, focusing on "microexpressions"-tiny muscle movements often too quick for normal perception. This led to his development of the Facial Action Coding System (FACS) in 1978, which identified about forty distinct facial muscle contractions called Action Units. Though FACS required extensive training and was labor-intensive to use, it created a standardized system for measuring facial expressions that would become foundational for machine learning applications.
As computer-based affect recognition developed, researchers created standardized image collections like the Cohn-Kanade dataset, where subjects performed 23 facial displays that FACS experts coded. Other collections followed, including the Karolinska Directed Emotional Faces database featuring extreme posed expressions matching Ekman's categories-exaggerated displays of surprise, joy, and fear that essentially created machine-readable emotion.
Despite Ekman's widespread influence, his work faced mounting criticism across disciplines. Anthropologist Margaret Mead challenged his binary view of emotions as either universal or not, arguing for a more nuanced position. Psychologists highlighted that fundamental questions remain unresolved, including whether facial expressions actually express emotions at all. Researchers Maria Gendron and Lisa Feldman Barrett warned specifically against AI applications of Ekman's theories, noting that detecting facial movements doesn't reliably indicate internal states.
Even as evidence points to their unreliability, companies seek new sources of facial imagery for profit. The narrow understanding of emotions-based on Ekman's limited set of anger, happiness, surprise, disgust, sadness, and fear-cannot capture the infinite universe of human feeling across cultures and contexts. This represents another instance of oversimplifying complex human experiences to make them computable and marketable.
7장
State: The Military Origins of AI and Its Surveillance Legacy
The Snowden archive reveals how intelligence agencies pioneered many AI techniques, showing a parallel AI sector developed in secrecy. Documents from 2013 describe programs like TREASUREMAP (designed to map the entire internet in real-time) and FOXACID (for compromising target computers), which resemble today's commercial surveillance systems but with fewer ethical constraints.
U.S. intelligence agencies have been major drivers of AI research since the 1950s, with military priorities of command, control, and surveillance profoundly shaping the field's development. The Defense Advanced Research Projects Agency (DARPA) became "the primary patron for the first twenty years of AI research," infusing battlefield-oriented concepts like target identification and anomaly detection into AI's foundational logics.
AI systems operate within complex multinational networks rather than pure national systems. Despite this reality, rhetoric positions AI development as a war between superpowers, particularly the U.S. and China. This framing serves to assert sovereign power over AI and reposition transnational tech companies within national boundaries.
The "Third Offset" strategy, championed by former Defense Secretary Ash Carter (2015-2017), sought to compensate for military disadvantages through technological superiority. Following nuclear weapons (First Offset) and advanced conventional weapons (Second Offset), this Third Offset would combine AI, computational warfare, and robotics. To achieve this, the Department of Defense needed to partner with tech companies that had already built the infrastructures and expertise the military lacked.
In April 2017, the Department of Defense announced Project Maven, an Algorithmic Warfare Cross-Functional Team designed to integrate AI and machine learning into military operations. Google won the first contract, using their TensorFlow infrastructure to analyze drone footage and detect objects and individuals. However, when Google employees discovered their work was being used for warfare purposes, over 3,100 signed a protest letter, forcing Google to withdraw from the project and release AI Principles prohibiting weapons development.
The relationship between the state and AI industry extends beyond military applications to local government functions, with militarized pattern detection and threat assessment technologies moving into municipal services. Palantir's point system creates a self-perpetuating cycle where individuals with high point values face heightened surveillance, increasing their likelihood of being stopped and further raising their point value. This machine learning approach leads to feedback loops where those in criminal justice databases face more scrutiny, deepening inequity while appearing objective through technological justification.
Despite the massive expansion of government AI contracts, little accountability exists for private vendors whose systems cause harm. Most states disclaim responsibility for problems created by AI systems they procure, arguing they cannot be responsible for systems they don't understand. This leaves commercial algorithmic systems influencing government decisions without meaningful accountability mechanisms.
Companies like Vigilant Solutions operate by taking surveillance tools that would require judicial oversight if operated by governments and turning them into private enterprises beyond constitutional privacy limits. Vigilant's automatic license plate recognition cameras create massive databases sold to police, private investigators, and companies. Similarly, Amazon's Ring doorbell cameras and Neighbors app create residential surveillance ecosystems that blur public-private boundaries.
At the heart of military targeting logic is the concept of the "signature"-patterns of behavior used to identify potential threats. During the Obama administration, the NSA's metadata surveillance program geolocated suspects for drone strikes based on device data rather than confirmed identity. As General Michael Hayden admitted, "We kill people based on metadata." This approach prioritizes correlation over precision, with devastating consequences-a Reprieve report revealed that drone strikes targeting 41 individuals resulted in 1,147 deaths.
This pattern recognition approach has expanded beyond military applications into refugee screening and social classification. During the 2015 Syrian refugee crisis, IBM developed an experimental "terrorist credit score" using harvested social media data and metadata to allegedly distinguish terrorists from refugees. These military and policing logics have merged with financialization, as credit-scoring models influence everything from loans to border crossings.
8장
Beyond Enchanted Determinism: Challenging AI's Power Structures
Artificial intelligence is not an objective, neutral computational technique making determinations without human direction. Rather, AI systems are deeply embedded in social, political, cultural, and economic contexts, shaped by humans and institutions to serve specific interests. They are designed to discriminate, amplify hierarchies, and encode narrow classifications. When deployed in social contexts like policing, courts, healthcare, and education, they reproduce and optimize existing structural inequalities.
Games have been AI's preferred testing ground since the 1950s because they offer closed worlds with defined parameters and clear victory conditions-unlike the messiness of real life. This approach stems from military-funded research that sought to simplify the world through mathematical formalism, emphasizing rationalization and prediction.
The epistemological flattening of complexity into clean signal creates what Crawford calls "enchanted determinism"-AI systems are portrayed as mystically powerful yet deterministically accurate. Deep learning's uninterpretability gives these systems an aura of being too complex to regulate and too powerful to refuse, obscuring power dynamics and closing off critical public scrutiny.
This enchanted determinism manifests in two mirrored narratives: tech utopianism that presents computational solutions as universally applicable, and tech dystopianism that blames algorithms as independent agents without examining their contexts. Both perspectives are ahistorical, locating power solely within technology while ignoring systemic forces like neoliberalism, austerity politics, racial inequality, and labor exploitation.
When AlphaGo defeats human champions, it's tempting to see otherworldly intelligence, but the reality is simpler: statistical analysis at scale. These systems can play millions of games and optimize for winning outcomes at speeds no human can match. Yet the ideology of Cartesian dualism persists-the fantasy that AI systems are disembodied brains producing knowledge independently from their creators, infrastructures, and material contexts.
Google's data center in The Dalles, Oregon offers a contrasting illustration of AI's reality-a massive facility using enough energy to power 82,000 homes, drawing on cheap electricity from the Columbia River. This infrastructure reminds us how much AI's expansion has been publicly subsidized: from defense funding and federal research to public utilities and tax breaks, not to mention the data and unpaid labor harvested from search engine users and social media participants.
How do we truly see AI's full life cycle and the power dynamics driving it? We must look beyond conventional maps to locate AI within a wider landscape. AI emerges from Bolivian salt lakes and Congolese mines, built from crowdworker-labeled datasets attempting to classify human identities. It guides drones over Yemen, directs immigration police, and modulates credit scores worldwide. Understanding AI requires a wide-angle, multiscalar perspective to contend with these overlapping regimes.
Should we simply democratize AI to serve justice rather than power? This appealing notion falters because AI's infrastructures inherently skew toward centralized control. As Audre Lorde reminds us, "the master's tools will never dismantle the master's house."
The technology sector's common response-signing AI ethics principles-falls short. These principles lack implementation mechanisms, enforcement, or accountability to the public. Unlike medicine or law, AI has no formal professional governance structure or ethical standards. Self-regulating frameworks allow companies to define "ethical AI" for the world while rarely facing consequences when violating their own principles.
To understand what's truly at stake requires focusing less on ethics and more on power. AI inevitably amplifies the forms of power it's deployed to optimize. Countering this means centering the interests of most-affected communities rather than glorifying founders, venture capitalists and technical visionaries. We should begin with the lived experiences of those disempowered, discriminated against, and harmed by AI systems.
The next era of critique must overturn the dogma of AI's inevitability. When AI expansion is seen as unstoppable, we can only patch together legal and technical restraints afterward. But what if we reverse this polarity and begin with commitment to a more just and sustainable world? How might we address interdependent issues of social, economic, and climate injustice? Where does technology serve that vision? And where should AI not be used because it undermines justice?
This is the basis for a renewed politics of refusal-opposing technological inevitability narratives. Rather than asking where AI will be applied merely because it can, we must question why it ought to be. By refusing systems that further inequity and violence, we challenge the structures of power that AI reinforces and create foundations for a different society.
