Blog · Review Essay · Last reviewed June 23, 2026

Atlas of AI and the Hidden Body of the Machine

Kate Crawford's Atlas of AI is a necessary corrective to the fantasy of artificial intelligence as clean cognition. It treats AI as a planetary system: mines, energy, warehouses, data extraction, classification schemes, affect recognition, state power, and corporate concentration arranged behind a seamless interface.

For this review, extraction is a governed transfer, not just a resource grab. Land, energy, water, labor, images, speech, records, classifications, and public authority are converted into machine capacity while the interface hides the terms of transfer: who paid, who consented, who can refuse, who owns the incident, and who can repair the damage.

The Book

Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence was published by Yale University Press in 2021. Yale's official page lists the book at 336 pages, with an ebook published April 6, 2021 and a paperback published August 16, 2022. The publisher presents the book as an account of AI's hidden costs across natural resources, labor, privacy, equality, and freedom.

Crawford's central move is to reject the usual image of AI as disembodied intelligence. The book asks what has to be extracted, sorted, labeled, optimized, surveilled, cooled, financed, and naturalized before a system can appear to answer as if it were a mind.

The book extends a research-art method Crawford had already developed with Vladan Joler in Anatomy of an AI System: take one polished interface, then follow it outward into labor, minerals, logistics, data, energy, disposal, and institutional power. Atlas of AI turns that anatomical method into a broader political map.

The key word is extraction, but the argument is sharper than saying technology companies take things. Extraction is an arrangement of dependency: the interface hides a chain of land use, labor discipline, data capture, institutional permission, and political risk. The useful review question is therefore not whether an AI tool seems impressive. It is what the tool needs in order to exist, who can refuse those needs, and who absorbs the damage when the system scales.

That frame is especially useful for current foundation-model systems because their apparent unity is a product effect. A chatbot, classifier, recommender, or agent may look like one service, but it is assembled from compute contracts, datasets, labels, moderation rules, human feedback, interface choices, cloud dependencies, benchmarks, and legal assumptions. Crawford's book teaches readers to ask for the hidden assembly, not just the visible answer, because the hidden assembly is where responsibility becomes enforceable or disappears.

Current Context

As of June 23, 2026, Crawford's extraction map is easier to test against public evidence. IEA now reports data-center electricity demand as an energy-policy question, DOE and Lawrence Berkeley National Laboratory quantify U.S. data-center load, NIST treats generative AI as a lifecycle risk-management problem, and the EU AI Act turns some high-risk AI uses into documentation, logging, oversight, monitoring, and incident-reporting obligations.

The current lesson is not that every model has the same footprint or that a single global number can judge a specific deployment. The lesson is that context cannot be treated as atmospheric. Training and inference mix, data-center region, power and water assumptions, labor suppliers, data rights, classification schemes, deployment sector, and appeal paths all change the risk profile. A serious AI adoption decision has to ask where the system touches the world before it asks how polished the answer sounds.

That makes the book a procurement test as much as a cultural critique. A school, hospital, benefits agency, newsroom, workplace, or civic archive should not ask only whether an AI service performs well in a demo. It should ask whether the service's compute siting, data provenance, labor chain, taxonomy, monitoring plan, and remedy path are specific enough to govern.

The Map

The title is exact. This is an atlas, not a manual. Crawford moves across domains that are usually discussed separately: the earth beneath the device, the warehouse behind automation, the dataset behind machine perception, the label taxonomy behind prediction, the face behind affect recognition, and the state behind surveillance.

Francesca Rossi's review in Artificial Intelligence summarizes the structure clearly: the book's chapters move through natural resources, human labor, personal data, classification, emotion recognition, and state power. Rossi also notes that Crawford's framing challenges the analogy between human and artificial minds by restoring embodiment, environment, and social relation to the analysis.

The result is a change in scale. AI is treated as a world-organizing system whose costs are displaced into places and people that the interface teaches users not to see. That is why this review belongs beside Feeding the Machine, the data-center governance essay, and the review of The Stack. Each asks what becomes politically possible when infrastructure is treated as atmosphere.

The atlas form also prevents a common audit failure. A model evaluation can tell you how a system behaves on a test. It cannot tell you whether the electricity was locally contestable, whether the labor chain was decent, whether the data rights were settled, whether the categories were legitimate, or whether public institutions became dependent on a private platform. Those questions require a map with separate layers for power, water, labor, data, taxonomy, deployment authority, and remedy.

Labor and Extraction

The strongest pages are about extraction in the broad sense. Minerals and energy make computation material. Warehouses and assembly lines show how automation often reorganizes human work before it replaces it. Data-labeling and content-classification labor keep the supposedly autonomous system fed, cleaned, and legible.

That point has become more visible since the book appeared. The International Energy Agency's 2026 update reported that global data-center electricity demand grew by 17 percent in 2025 and projected roughly a doubling from 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030, with AI-focused data centers growing faster than the sector as a whole. In the United States, Lawrence Berkeley National Laboratory's 2024 report estimated that data centers used 176 terawatt-hours in 2023, about 4.4 percent of total U.S. electricity consumption, and modeled a 2028 range of 325 to 580 terawatt-hours.

Those numbers do not make every model illegitimate. They do make AI deployment an industrial-policy problem, not only a software release. Grid interconnection, water use, local permitting, tax incentives, labor standards, and emergency power are part of the system's real design. A public institution buying an AI service should be able to ask for a supply-chain ledger: compute provider, hosting region, labor supply, data rights, moderation pipeline, energy and water assumptions, failure reporting, and refusal or appeal channels for affected people. The same operational demand appears in the site's AI bill-of-materials proposal and data-sheet supply-chain checklist.

The labor side needs the same discipline as the energy side. If a system relies on annotation, content moderation, red teaming, preference ranking, data cleaning, voice work, warehouse routing, or customer-support triage, that work is not a footnote to model quality. It is part of the production system. Calling the finished product automatic can erase the workers who made the system measurable, safe enough to sell, and cheap enough to scale.

Classification as World-Making

Atlas of AI is especially valuable on classification. A model does not merely learn from the world. It learns from a world already cut into categories. Those categories carry histories: race, gender, normality, risk, productivity, emotion, threat, and usefulness.

Classification becomes a political act when it is embedded in systems that allocate attention, suspicion, services, policing, credit, employment, or care. A label can look technical while smuggling in an institution's preferred theory of the person.

Crawford's work with artist Trevor Paglen on Excavating AI made the argument concrete by examining ImageNet's person categories and the judgments built into a supposedly neutral training set. Their project documented derogatory, racialized, gendered, and insulting categories. ImageNet's own September 2019 update said the full dataset contained 2,832 people categories under the person subtree, identified three sources of downstream harm, and described a planned removal of 600,040 images from unsafe synsets. The episode is the book's thesis in miniature: a taxonomy assembled cheaply and treated as neutral can quietly encode judgments about who counts as what, then move those judgments into models, benchmarks, and downstream institutions.

Once a category enters a database, the database can train a model; once a model predicts through that category, institutions can reorganize around the prediction; once people adapt to the institution, the category begins to look like evidence that it was natural all along. This is the same problem examined from different angles in Sorting Things Out, Algorithms of Oppression, and Weapons of Math Destruction.

The governance lesson is that classification must be audited before it becomes infrastructure. A data sheet should name the taxonomy, source institutions, labeling process, contested categories, known exclusions, intended use, prohibited uses, affected groups, and appeal path. A model card or system card should not hide those choices behind aggregate accuracy. The question is not only whether the model predicts a label correctly; it is whether the label should govern anyone at all. If the label concerns identity, ability, emotion, risk, productivity, threat, fraud, or eligibility, the burden should be on the deployer to justify the category before the system touches people.

State Power and Surveillance

Crawford also links AI to state power. The point is not only that governments use AI systems. It is that techniques built around classification, detection, targeting, and prediction fit comfortably inside institutions that already want more visibility, faster intervention, and cleaner administrative files.

By June 2026, that concern had become a live regulatory category. The EU AI Act, Regulation (EU) 2024/1689, restricts several biometric and emotion-recognition uses and classifies many systems in employment, education, law enforcement, migration, and access to essential services as high risk. Its high-risk obligations emphasize documentation, logging, human oversight, monitoring, and incident reporting as the law phases in. NIST's AI Risk Management Framework and its generative-AI profile point in the same practical direction: risk management must cover lifecycle documentation, measurement, governance, and use context, not just model performance.

The safety implication is direct. State and public-service use of AI should require a named purpose, a legal basis, tested error rates for the affected population, logging, human accountability, appeal channels, procurement transparency, and an exit plan. A system that classifies bodies for policing, benefits, hiring, or education is not an oracle. It is an administrative instrument. The related review of Automating Inequality shows what happens when that instrument disappears into welfare bureaucracy.

The most dangerous deployments are therefore not only inaccurate systems. They are systems that make extraction administrative: a face becomes a risk signal, a household becomes a benefits case, a worker becomes a productivity trace, a student becomes a behavioral category, and a neighborhood becomes a predictive policing surface. The harm is not just bad prediction. It is the power to make a contested category hard to refuse.

The AI-Age Reading

Read in 2026, Atlas of AI is less a warning from the past than a field guide for current foundation-model governance. The generative-AI boom has made Crawford's map easier to verify: energy agencies now track data-center load, training-data provenance is a procurement and legal-risk question, and regulators are turning documentation into a legal duty.

The strongest use of the book is operational. Before adopting an AI system, an institution should ask for a supply-chain account: compute provider, training and tuning data sources, labor conditions for labeling and moderation, safety evaluations, incident history, known limitations, content provenance support, environmental estimates, and rights of appeal for people affected by outputs. Model cards, system cards, provenance standards, impact assessments, and public-interest procurement rules are not bureaucratic decoration. They are the paperwork that makes hidden dependencies governable. See the site's notes on model cards and system cards, content provenance, algorithmic impact assessments, and public compute for the constructive side of that program.

The book also changes how to read safety claims. A vendor may report refusal rates, benchmark scores, red-team results, and policy compliance. Those are useful, but they are not the whole safety case. If the system depends on opaque data rights, hidden annotation labor, concentrated cloud capacity, uninspected classification schemes, or a public agency with no meaningful appeal channel, the safety case is missing part of the system.

Governance and Safety

Read on June 23, 2026, Crawford's extraction lens translates into a practical governance standard: make the hidden chain inspectable before deployment, not after scandal. NIST's AI Risk Management Framework remains voluntary, but it frames risk management across design, development, use, and evaluation; NIST AI 600-1 applies that risk-management approach to generative AI. The EU AI Act gives the same idea legal teeth for high-risk systems by requiring instructions for use, human oversight, logging support, monitoring, deployer duties, and serious-incident reporting where the Regulation applies.

The concrete artifact is an extraction ledger. Before deployment, the buyer should be able to see the system's material inputs, compute region, energy and water assumptions, data sources, rights claims, labor vendors, moderation and human-feedback work, taxonomy, model and evaluation versions, deployment purpose, affected population, refusal path, appeal path, incident owner, update policy, and shutdown criteria. That ledger is not a sustainability appendix. It is the deployment's basic safety record, because hidden dependencies are where liability, rights violations, and operational failure usually hide.

For material infrastructure, the safety question is whether compute demand is locally legible. The IEA's 2026 update and the Lawrence Berkeley National Laboratory report make data-center load too large to treat as a background cloud cost. Buyers and public agencies should ask where compute is hosted, what energy and water assumptions support the service, what backup and resilience plans exist, and whether public infrastructure or ratepayers are subsidizing private model capacity.

For data and labor, the safety question is provenance. Procurement should require records for training data, fine-tuning data, evaluation sets, labeling and moderation vendors, human-feedback work, rights claims, opt-out or deletion duties, and known gaps. A model that cannot describe its data and labor dependencies is not simply opaque. It is asking the buyer to inherit unknown social, legal, and operational risk.

For classification systems, the safety question is contestability. Any deployment that scores, sorts, detects, infers emotion, identifies bodies, ranks workers, allocates services, or flags risk should have a named purpose, tested group performance, logging, human authority to override, notice where required, appeal channels, incident reporting, and a withdrawal plan. The EU AI Act's high-risk categories in biometrics, education, employment, essential services, law enforcement, migration, and justice show why taxonomy is not merely technical metadata.

The decision standard should be explicit. A system can be allowed, conditioned, delayed, redesigned, locally negotiated, publicly owned, independently audited, or rejected. The extraction ledger gives reasons for that decision. Missing provenance, no accountable incident owner, no appeal route, no lawful or proportionate purpose, unverifiable labor claims, or material infrastructure costs shifted onto the public should count as deployment risks, not public-relations problems.

The governance mistake is to reduce extraction to a moral mood. The operational response is a dated supply-chain map: materials, compute, data, labor, categories, vendors, model updates, deployment context, affected people, and remedies. Without that map, safety work becomes a surface exercise over a system whose real dependencies remain politically unavailable.

Where the Book Needs Friction

The book's weakness is the other side of its strength. Its broad critique sometimes compresses AI, machine learning, platforms, datafication, and automation into one political object. Rossi's Artificial Intelligence review makes a useful technical objection: AI as a scientific field is wider than machine learning, including areas such as search, planning, scheduling, optimization, and knowledge representation.

That distinction matters. If every computational system is treated as the same kind of AI system, the analysis can lose precision. A procurement rule, a vision model, a warehouse scanner, a recommender system, a medical triage model, and a chatbot may share political conditions without sharing the same technical failure modes.

Yuchao Zhao's 2025 review in Global Media and China is also helpful here. It praises the book's interdisciplinary force while noting that some datafication arguments are not new. That is fair. Crawford's originality is less in discovering each component than in forcing them onto the same map. The right response is not to soften the extraction lens, but to make it more precise: which resource, which institution, which labor relation, which data claim, which classification, which feedback loop, which failure mode?

Another limit is remedy. Naming extraction does not itself decide when a system should be banned, delayed, redesigned, locally negotiated, publicly owned, independently audited, or allowed with conditions. That judgment needs sector-specific evidence. A community-facing benefits model, a research classifier, a medical imaging system, and a large language model embedded in an office suite may all involve extraction, but the safety controls and public remedies will differ. The point is not to flatten every system into one indictment. It is to make each system show its dependencies before it receives institutional authority.

What This Changes

Atlas of AI is a book about the machine's hidden body.

Artificial intelligence often presents itself as answer, assistant, coworker, companion, or neutral infrastructure. Crawford keeps asking what vanished in order for that presentation to work. The mine vanishes into the chip. The worker vanishes into automation. The labeler vanishes into the dataset. The dataset vanishes into fluency. The server farm vanishes into the word "cloud." The institution vanishes into the interface.

That disappearance is politically dangerous because it turns dependency into atmosphere. When users experience AI as a clean surface, they can forget the supply chains, labor regimes, data claims, classification choices, and surveillance institutions underneath it.

The practical lesson is to audit AI from the ground up. Ask where the materials come from, where compute is hosted, who performs the hidden work, what data was taken, which categories were imposed, what institutions gain power, and who can refuse, appeal, inspect, or exit. Without that map, AI ethics becomes a set of decorative principles floating above the system that actually governs.

The sharper institutional lesson is to make disappearance a refusal condition. If the vendor cannot disclose enough about compute, data provenance, labor chain, model updates, classification logic, monitoring, incident reporting, and appeal paths for the buyer to govern the system, the system should not be treated as clean infrastructure. Opacity is not a neutral business constraint when the product reorganizes public memory, labor, services, or suspicion.

Source Discipline

This review separates bibliographic, interpretive, operational, legal, and implementation claims. Yale University Press verifies the book's metadata and publisher framing. Rossi and Zhao supply scholarly review context. Crawford and Joler's Anatomy of an AI System and Crawford and Paglen's Excavating AI are treated as primary interpretive projects, while ImageNet's own update is used for the dataset-maintenance claim. Energy-demand figures come from IEA, DOE, and Lawrence Berkeley National Laboratory sources. Regulatory claims come from EUR-Lex and NIST. Dataset-documentation claims use Datasheets for Datasets and related documentation work. None of those sources proves that every AI system has the same footprint or the same remedy.

Claims about extraction should name the layer: mineral supply, electricity, water, labor, data provenance, classification scheme, biometric use, workplace management, public-sector deployment, cloud dependency, or public remedy. "AI is extractive" is a weak slogan unless the reader can see what is being extracted, by whom, under what permission, with what public risk, and with what route for contesting or repairing the harm.

Source discipline also means treating corporate sustainability pages, product announcements, and benchmark reports as limited evidence. They may document commitments or measured behavior, but they do not replace utility filings, standards, legal duties, worker testimony, procurement records, independent evaluations, dataset documentation, or incident reports. If a vendor cannot distinguish a safety claim from a marketing claim, the buyer should treat the claim as unverified.

Sources

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

Buy on Amazon Browse Books

Return to Blog · Return to Books