Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence
Power, Politics, and the Planetary Costs of Artificial Intelligence
The question Crawford asks is embarrassingly simple: before you celebrate what AI can do, have you looked at where it came from? *Atlas of AI* is a book about the material substrate beneath every language model, image classifier, and recommendation engine — the lithium dragged out of Nevada desert, the Amazon warehouse workers whose injury rates are carefully not disclosed, the facial recognition datasets scraped without consent from Flickr and mugshot databases. It's a book that insists on making visible what Silicon Valley's marketing budget works hard to obscure.
The argument is organized as an actual atlas — each chapter a thematic map of AI's true territory. The "Earth" chapter is the strongest: Crawford visits Silver Peak, Nevada, where lithium is pumped from underground brine lakes to feed the batteries that feed the computation. The contrast she draws — between the clean product experience of an iPhone and the dusty, underpaid, ecologically depleted reality behind it — is legitimate and well-executed. The "Labor" chapter similarly traces how platforms like Amazon Mechanical Turk and content moderation operations essentially depend on a global invisible workforce doing the work that makes machine learning look like it works. This is where Crawford is at her most useful: documenting the gap between how AI presents itself and what it actually requires.
To understand how AI is fundamentally political, we need to go beyond neural nets and statistical pattern recognition to instead ask what is being optimized, and for whom, and who gets to decide.
— Crawford, *Atlas of AI*, Introduction, p. 9
Where the book earns more skepticism is in the political conclusions it builds toward. Crawford argues AI is fundamentally a "registry of power," a tool through which the existing order reproduces and amplifies itself. This is often true, and the historical evidence she marshals — from phrenology's influence on early classification systems to the military origins of satellite imagery — is genuinely illuminating. But the analysis tips toward treating power concentration as AI's inevitable destiny rather than its current configuration. The technology that enables facial recognition for state surveillance is the same technology that powers tools used by dissidents and journalists. Crawford notes this in passing; she doesn't sit with it. You come away with a coherent critique of AI as it actually operates under the incentives of 2020s capitalism, but the implicit prescription — be more suspicious, demand accountability, slow down — doesn't quite square with the scale of what's actually happening. The question of whether the costs she documents are inherent to the technology or inherent to *this moment* in its development goes mostly unasked.
artificial intelligence is now a player in the shaping of knowledge, communication, and power
— Crawford, *Atlas of AI*, Introduction, p. 19
For a technical reader, the book is deliberately not a how-it-works manual — Crawford is explicit that she's writing political economy, not computer science. Cathy O'Neil's *Weapons of Math Destruction* or Gary Marcus and Ernest Davis's *Rebooting AI* will do more for you if you want to understand the machinery. What Crawford does well is what no technical book does: she forces you to see AI as infrastructure, with all the supply chains, labor relations, environmental footprints, and governance questions that infrastructure implies.
plausible deniability for any exploitative practices that drive their profits
— Crawford, *Atlas of AI*, ch. 1 "Earth", p. 35
Worth reading because the costs are real, even if you think the benefits outweigh them. Especially worth reading if you've never considered that your query to a language model ends, many steps back, at a lithium mine.
Read the longer summary
Crawford’s central claim
The argument is direct. Artificial intelligence does not really earn either word of its name, and treating it as if it does obscures what AI actually is: an industrial system that pulls minerals out of the ground, labor out of low-wage workers, and data out of anyone with a phone, then channels the output into the hands of a small number of companies and states. Crawford calls AI “a registry of power” (p. 8), and the whole book is a sustained argument for that phrase.
The framing device is the atlas. Each chapter is a different map of the same territory: Earth, Labor, Data, Classification, Affect, State. The book does not try to give you a technical primer on how a transformer works. Crawford is an interdisciplinary scholar (Microsoft Research, co-founder of the AI Now Institute at NYU, visiting chair at the École Normale Supérieure), and her tools come from anthropology and political economy, not engineering. If you want to know how gradient descent works, this is not your book. If you want to know what has to be dug out of the ground before a single GPU can train a model, it is.
Crawford is explicit about the politics. She wants the reader to see AI as “politics by other means” (p. 20), a way for the people who build and own these systems to encode their preferences into infrastructure the rest of us then have to live inside. That framing tells you what the book is and isn’t. It is a polemic in the older sense: a sustained, well-researched argument with a thesis. It is not a balanced primer.
The material chapter is the best one
The chapter called “Earth” is where the book is strongest, and probably where it justifies its existence. Crawford drives to Silver Peak, Nevada, an old silver-mining town that now sits over an underground reservoir of lithium brine, and she walks the reader through what it takes to make a battery, and therefore what it takes to make a phone, a data center, a car-with-AI-inside. She does the same for rare earths in Inner Mongolia and for latex tapped from rubber trees in Malaysia. (A useful reminder that even the keyboard on your laptop has a plantation history.)
The point she is hammering, and it is a fair one, is that the cloud is a misleading word. Data centers consume electricity and fresh water on the scale of small countries. The phrase she uses, that complex supply chains hand tech companies “plausible deniability for any exploitative practices that drive their profits” (p. 35), is one of the book’s better lines, because it identifies a specific mechanism rather than gesturing at vibes. When a labor abuse happens four contractors deep, the brand at the top genuinely does not know about it, and arguably has structured itself not to know.
This chapter pairs well with the next one, “Labor,” which is mostly about the hidden human work that gets papered over by the word automation. Mechanical Turk workers tagging images for cents per task. Content moderators traumatized by the worst of the internet so the rest of us don’t have to see it. Warehouse workers whose movements are timed to the second by the same logistics software that delivered the book to your door. Crawford’s argument here lines up with reporting that has only intensified since the book came out. The data-labor supply chain in Kenya and the Philippines that powers the safety filters of every major chatbot is now one of the better-documented stories in tech, and Crawford got there early.
We think this material, the mines and the water and the labelers, is the part of Atlas of AI that will hold up longest. It is also the part that is hardest to argue with, because it is mostly empirical. Either lithium has to come from somewhere or it doesn’t.
Data, classification, and the eugenics throughline
The middle of the book gets more contested. “Data” opens at a National Institute of Standards and Technology archive, with mugshots from a forensic database that Crawford uses to draw a line from nineteenth-century anthropometry (the calipers-and-skull-measuring tradition) to today’s facial-recognition training sets. The argument is that the AI industry inherited a habit of treating people as datapoints to be classified, scored, and ranked, and that this habit is older and more politically loaded than the field tends to acknowledge.
There is real substance here. The ImageNet labeling controversy, where researchers found racial slurs and sexually explicit categories baked into one of the foundational image datasets in computer vision, is a story Crawford tells well. So is the broader point that machine-learning datasets get assembled, often by graduate students under deadline, with categories whose history nobody bothered to audit, and then those categories get frozen into models that millions of products are built on. That is a real engineering failure mode. We have seen it firsthand.
The chapter that follows, “Classification,” extends this. The act of putting people into bins is never neutral, the bins always reflect someone’s worldview, and AI systems scale that worldview to billions of decisions. We agree with the descriptive claim. The question is how far you can push the historical lineage. Crawford sometimes writes as if there is a more or less unbroken thread from Francis Galton to TensorFlow, which makes for a tidy story but is doing a lot of work. Galton was a eugenicist. Galton also invented standard deviation. Statistics is not a moral category. The fact that a gradient-boosted tree will give you correlated outcomes by race in a country with correlated outcomes by race is not a fact about the gradient-boosted tree; it is a fact about the country. Crawford knows this. The book sometimes elides it for rhetorical lift.
Where the framing strains
The chapter called “Affect” is on the weakest ground. Crawford goes hard at companies that claim to read emotion from facial expressions, including Affectiva, the now-rebranded HireVue products, and various security-state contractors. She has a strong empirical case. The underlying psychology, drawn mostly from Paul Ekman’s claim of universal basic emotions, is much more contested than the products imply, and the products themselves perform poorly on faces from outside the populations the training data was drawn from. Calling out vendor hype here is fair game.
But the chapter sometimes slides from “this product doesn’t do what it says” into “the entire idea of inferring affective state from observable behavior is a category error.” That is a much bigger claim, and one that is harder to sustain when the same volume insists, correctly, that AI systems are very good at detecting patterns humans miss. The book wants emotion detection to be uniquely fraudulent in a way that, say, fraud detection is not, and the line is harder to draw than Crawford lets on.
The bigger issue, and this is where we have to push back, is the load-bearing claim that AI is supposedly neither of the two things its name says it is. That was a defensible rhetorical move in 2021. It reads very differently after late 2022. Whatever GPT-4 and its successors are, the dismissive framing of AI as “just” pattern-matching on human-generated data has been overtaken by systems that produce novel proofs, novel code, and novel arguments. Including, we’d note, arguments that critique their own training data in ways their developers did not anticipate. Crawford’s point about extraction can be true and the dismissive framing about machine intelligence can be wrong at the same time. The book is weaker for tying them together.
A volume published in 2021 cannot be blamed for not predicting the LLM era. But it can be blamed for the rhetorical moves that the LLM era made obviously brittle, and “AI isn’t really intelligent anyway” was already a thin claim when the book shipped.
The extraction frame applied to everything
Crawford’s central metaphor is extraction. Minerals out of the earth, labor out of workers, data out of users, value out of all of the above and into a few balance sheets in San Francisco and Seattle. The metaphor does a lot of work, and in the Earth and Labor chapters it does honest work. The lithium really is being extracted. The workers really are being underpaid. The supply chain really is opaque.
Where the metaphor strains is when it becomes the only frame. Every interaction with a tech company becomes extraction. Every classification system becomes a colonial project. Every dataset becomes loot. At that point the analysis stops doing the thing analysis is supposed to do, which is distinguish cases. A free Gmail account is a different kind of bargain than a cobalt mine in the Congo, even if you can describe both as extractive in a long enough chain of metaphors. The book sometimes treats the chain as the point. If everything is extraction, the word stops carrying weight.
The other underdeveloped angle is the alternative. Atlas of AI is very clear about what it does not want: concentrated, opaque, environmentally costly, politically unaccountable AI. It is much vaguer about what it does want. There are gestures toward data justice, collective governance, and the AI Now Institute’s policy work. There is no serious engagement with the case that the same supply chains carrying lithium to data centers also carry insulin to clinics and electric vehicles to driveways, and that decoupling is not free. The book treats the costs as nearly all on one side of the ledger. That is a polemic’s prerogative, but the reader should know it’s the genre.
What the reader still needs
Crawford gets you the political and material picture. She does not get you the technical picture, and she is honest enough not to pretend otherwise. If you want to argue with the AI industry on its own ground, you also need to know how it works on its own terms: how a model gets trained, what backpropagation actually does, why scale matters. Lauren Goodlad’s review for Critical Inquiry points readers toward Cathy O’Neil, Meredith Broussard, Gary Marcus and Ernest Davis for the technical side. That is sound advice. Atlas of AI on its own will not let you debug a model. It will let you ask why the model exists and who gains.
You also need a counterpoint. This is from one side of a real argument, and reading only one side is how you become the kind of person Crawford is implicitly writing against: confident about a topic you have only heard one framing of. Pair it with something from the other camp. Andrew McAfee’s More from Less makes the harder version of the case that decoupling resource use from output is real and ongoing. Read both and make your own decision about which mechanism dominates.
Who should read it
Engineers and product people who are building on top of these systems and have never thought about where the lithium comes from, or how the labels in their training set were produced, or what happens to a content moderator at the end of an eight-hour shift. The book will not change your day-to-day work, but it will change what you notice. That is worth a weekend.
Policy people who want a single book to anchor a critique of concentrated AI power. Atlas of AI is well-cited, well-written, and has the institutional credibility (Yale University Press, Microsoft Research, AI Now Institute) to be quoted in a memo without raising eyebrows. It will not give you the technical depth, but it will give you the frame.
We’d be more cautious about handing it to someone forming their first impression of AI. Read alone, the book gives you an unrelieved indictment. We think the indictment is mostly right on materials and labor, partly right on data and classification, and overstretched on intelligence and affect. A reader coming in fresh deserves the chance to weigh the strong parts against the weak ones, and Crawford is too good a stylist for the weak parts to read as weak on the first pass. Read it second, after you have a baseline. Read it with people who will argue back.