Kate Crawford, the Atlas of AI and what the press told us

"AI is neither artificial nor intelligent." With that line Kate Crawford opens her Atlas of AI (Yale University Press, 2021). And with that line, on its own, a good part of the Spanish-language press stopped. The quote runs around the networks like a slogan; the book that holds it up almost nobody has read in full.

I'll confess it up front, because the same happened to me. When I started writing here I defended something close to that cover line: that the name that won, "artificial intelligence", is so off the mark that it says more about our desire than about the thing. The machine couldn't care less what we call it. The problem is ours. We don't entirely know what it is and even so we've handed it the house keys and let it look in the fridge, assuming it won't do worse than us. What I hadn't done, what almost no one who repeats the line has done, is travel to the places where that thing weighs, sweats and rips minerals out of the ground. Crawford did go. That's why it's worth reading her in full, and not by the cover.

The uncomfortable thesis

The book places AI where it really lives. In the lithium brines of the Atacama salt flat. In warehouses timed to the second. In server farms drinking from aquifers in areas that were already short of water. What moves a prompt — the instruction typed into a chat — is matter, and that's exactly the idea the general press walked past: it underlined the aphorism and left the atlas out.

Crawford argues that AI works as an extractive industry. Like mining or trawler fishing, it withdraws resources from the world: minerals, energy, water, data, human labour. The metaphor of "the cloud" does the rest of the dirty work, because it erases that materiality and shifts the costs to someone paying far from the screen. Pressing enter looks free. It isn't free for everyone equally.

Her authority doesn't come from armchair activism. It comes from having walked the physical chains and documented them with names, places and dates, from a Chilean mine to the query someone types in Madrid. Hence the discomfort, too: it's hard to reduce to a headline a book built by walking through warehouses, government archives and police stations. What's more than a decade of fieldwork, spread across six thematic chapters, doesn't let itself be summed up in a tweet. Though the tweet is the only thing that survived.

Atacama, cobalt and the water nobody sees

The first chapter, Earth, opens with a trip to the Atacama salt flat. Crawford documents how lithium extraction for batteries presses on the groundwater and the ways of life of the Atacameño communities of the high plateau. Chilean lithium feeds batteries that feed the devices that hold up the infrastructure AI runs on. The chain isn't an allegory: it's made of roads, brines and water balances you can measure.

Then comes Congolese cobalt, the rare earths processed mostly in China, the purified silicon, the aluminium. Every component of a server has a geographical origin, a human cost and an environmental cost that don't show up on the user's bill. The Spanish press, when it mentions it, treats it as colour. And it mentions it rarely.

And then there's water, which is the easiest example to check because the companies themselves publish it. In 2023 Google declared a consumption of around 23 billion litres of water in its data centres, and Microsoft put its own at around 1.7 billion gallons. It's worth being honest about what that number says and what it doesn't: it's the total consumption of those companies' data-centre infrastructure, cooling included, not the isolated bill of training one specific model. But the direction points where Crawford was looking in 2021. The cloud competes for real aquifers in real municipalities. When someone drops the word "cloud" in a meeting, the honest question is which aquifer they're talking about.

The work that holds AI up and doesn't appear in the credits

There's a visible layer of work in this industry: the doctorates, the six-figure-salary engineers, the press conferences. Crawford devotes her harshest chapter, Labor, to the other layer. The one that's not in the photo.

There are the warehouse workers whose pace is set by an algorithm that times every gesture to squeeze the packages-per-hour rate, with effects on the body she tracks in her fieldwork. And there is, above all, the army of annotators and moderators who clean up what the models mustn't repeat. The best-documented case was broken by Time magazine in January 2023: workers in Kenya, hired by the firm Sama for OpenAI, were paid between 1.32 and 2 dollars an hour to label and filter text and images with descriptions of abuse, violence and explicit sexual content, so that ChatGPT and its relatives wouldn't reproduce them. It wasn't "one to three dollars" nor a figure rounded to a headline's taste. It was those numbers, in that contract, over that material.

There's a third, more diffuse variant: platform-mediated work, where a distribution of fees and metrics decided by software takes the place of the old-fashioned employment contract and can throw you out with no one to appeal to. Different layers, the same mechanics. The end user gets a clean, fast answer, as if out of nowhere. It comes from there.

Who decides the categories

The Classification chapter goes after something more elusive than scandal. Crawford asks how labels get assigned to the large datasets the models are trained on, and why those labels, far from neutral, carry inside them a way of looking at the world.

Her example is ImageNet, the dataset that in 2009 laid the foundations of modern computer vision and gathers around fourteen million images across thousands of categories. Most are harmless — "dog", "car", "table" — but not all.

In September 2019, Crawford and the artist Trevor Paglen launched ImageNet Roulette, a tool that showed which categories ImageNet used to classify the face of whoever uploaded their photo. Out came insulting, racist labels applied to real people who had never given their consent. After the fallout, the ImageNet team announced the removal of around 600,000 images from its person category and flagged about fifteen hundred more categories as problematic, between unsafe and sensitive. What the press release didn't underline: the models already trained on those images kept running in production.

Crawford's lesson goes deeper than the specific episode. Any classification system is a political decision. Deciding which things go together, which categories exist and which don't, what name each drawer gets — all of that fixes a worldview and disguises it as natural order. Labelling people by race, ethnicity or gender, she writes, resembles more than we'd like the old mania for measuring skulls. And the models inherit that load without knowing it: the dataset's biases, the gaze of whoever labelled, the gaps where there was no data. Then they spit out decisions about credit, admissions or surveillance with the impeccable air of statistics — which was exactly what anthropometry sought to lend to phrenology.

The State isn't outside. It's co-author

In State, Crawford follows the trail of the public money that holds up much of AI's advance in the United States. DARPA funded for decades research lines that led to today's neural networks, and the border between civilian and military AI has been honed until it almost disappeared.

What in 2021 she posed as a tendency has become a headline, and here I note it as my own dated observation, not as a quote from the book. In 2024 Anthropic teamed up with Palantir and Amazon Web Services to bring its Claude models into the US government's classified intelligence and defence environments. And in July 2025 the Department of Defense awarded contracts of up to 200 million dollars each to OpenAI, Anthropic, Google and xAI to develop AI capabilities for national-security uses. Crawford described the pattern before these contracts existed; the pattern hasn't proven her wrong.

The chapter's argument isn't catalogue pacifism, but something more structural. The AI we have wasn't born of the invisible hand nor the lone genius in the garage. It grew on sustained state investment, with state priorities and direct military applications. That dislodges two narratives in one stroke. To Silicon Valley, which boasts of innovating despite the government, it's a reminder that the government was at the cradle. And to whoever trusts that "regulation will set limits" it puts an ugly problem in front of them: when the regulator is at once the investor and the customer, the conflict of interest isn't an accident of the system, it's its design.

Anatomy of an AI System, the map before the book

Three years before the Atlas, in September 2018, Crawford and the researcher Vladan Joler published Anatomy of an AI System, a project of the AI Now Institute and SHARE Lab. It's still the best entry door to all of this because it is, literally, a map.

A huge diagram laying out what it takes for an Amazon Echo to answer when someone asks it the time: minerals ripped from the ground, components made on the other side of the planet, ships, lorries, data centres, people annotating data far from the cameras, electrical grids with their particular mix of sources, cooling water. All on a single sheet, freely available at anatomyof.ai.

It works because it makes the invisible visible without asking the viewer to read an essay. Just looking at the map is enough to understand that an "Alexa, what time is it?" sets a planetary chain in motion. AI is neither light nor ethereal, and it isn't "in the cloud" in any literal sense. It's in the ground, in the water and in the working days of people with names.

Why this book doesn't enter the Spanish debate

Before going on it's worth defusing a falsehood doing the rounds: that it isn't translated. It is. Atlas de IA came out in Spanish from NED Ediciones in 2023, translated by Francisco Díaz Klaassen. It exists, you can buy it. The real question is another: why a book that weighs so much in the Anglo debate barely shows up in the Spanish one except as decorative quote.

I have three suspicions, and I give them for what they are, my suspicions. The first is that the book doesn't fit the "good AI versus bad AI" scheme. It forces you to think across several scales at once — global, regional, local — about material chains, classification politics, the history of labour. That's effort, not headline, and the headline always reaches the newsroom first.

The second suspicion is that it lands its blows without checking the ID card. It hits Silicon Valley techno-utopianism and it hits the left that celebrates innovation without looking at the bill; it points at governments of one stripe and the other. A book that plays for no team is no use to any team trying to be right, and almost everything published about AI is published to be right.

The third is the one that interests me most. Translating its theses into Spanish everyday life would demand fieldwork here: lithium in Extremadura, data-centre projects in Aragón, water stress in specific municipalities, annotators in Latin America working under contracts signed in Spanish. That local connection is possible. But asking for the trip, and not just the review, is asking a lot of a newsroom in a hurry.

What breaks when you read it in full

After reading it, there's one word that stops being any use: "cloud". Once you've got the map in your head, the term sounds like an anaesthetic, a comfortable metaphor that exists precisely so you don't ask about the aquifer or the warehouse.

The word "bias", understood as an abstract laboratory category, also comes up short. You start asking where the specific dataset comes from, who labelled it, under what conditions, by what criterion. Bias turns into material history, with geography and with names, and you can see its seams.

And the political question mutates, which was Crawford's real target. If AI is an extractive industry with a planetary chain and a defined business model, switching from only asking "how do we regulate it" to also asking who pays the cost and who keeps the value isn't a rhetorical nuance. It's changing the whole conversation. The cover line, the one everyone repeats, was just the door. The book is inside, and inside it's colder.

Definitions

Atlas of AI. A book by Kate Crawford published by Yale University Press in 2021 and translated into Spanish by NED Ediciones in 2023 under the title Atlas de IA. It's organised into six thematic chapters — Earth, Labor, Data, Classification, Affect and State — plus a conclusion titled Power and a final coda, Space.

Extractive industry. An activity whose operation consists of withdrawing non-renewable or costly-to-regenerate resources, such as mining, hydrocarbons or trawler fishing. Crawford transfers the term to AI to underline that its functioning depends on extracting minerals, energy, water, data and labour.

Prompt. The text instruction a user writes to ask something of a conversational AI system.

ImageNet. A dataset of around fourteen million labelled images, the basis of modern computer vision since 2009. In 2019 those responsible announced the removal of around 600,000 images from its person category after the complaint by Crawford and Paglen.

Anatomy of an AI System. A visual project by Kate Crawford and Vladan Joler published in September 2018, mapping the material and human chains needed to make and operate an Amazon Echo. Available at anatomyof.ai.

AI Now Institute. A centre for critical research on AI founded in 2017 at New York University by Kate Crawford and Meredith Whittaker.

References

Crawford, K. — Atlas of AI: Power, Politics and the Planetary Costs of Artificial Intelligence (Yale University Press, 2021). Spanish edition: Atlas de IA, trans. Francisco Díaz Klaassen (NED Ediciones, 2023). The source for the six-chapter structure and the thesis of AI as an extractive industry.

Crawford, K. & Paglen, T. — ImageNet Roulette and the essay Excavating AI (online project, September 2019). On the classification of people in ImageNet and the subsequent removal of around 600,000 images.

The Art Newspaper / Artsy — coverage of the removal of around 600,000 images from ImageNet's person category and the flagging of about fifteen hundred categories as problematic, September 2019.

Crawford, K. & Joler, V. — Anatomy of an AI System (AI Now Institute and SHARE Lab, September 2018, anatomyof.ai). A visual map of the Amazon Echo's material chains.

Perrigo, B. — "OpenAI Used Kenyan Workers on Less Than \$2 Per Hour to Make ChatGPT Less Toxic", Time, January 2023 (time.com/6247678/openai-chatgpt-kenya-workers). The source for the wage figures (between 1.32 and 2 dollars an hour) of the Sama annotators for OpenAI.

Corporate environmental reports — water consumption of Google's data centres (≈23 billion litres in 2023) and Microsoft's (≈1.7 billion gallons in 2023), referring to the total consumption of their infrastructure.

CNBC and Breaking Defense — contracts of up to 200 million dollars from the US Department of Defense with OpenAI, Anthropic, Google and xAI, July 2025 (cnbc.com/2025/07/14/anthropic-google-openai-xai-granted-up-to-200-million-from-dod.html).

TechCrunch / Palantir Investor Relations — Anthropic's partnership with Palantir and Amazon Web Services to bring the Claude models into the US government's classified intelligence and defence environments, November 2024.

ImageNet — Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K. & Fei-Fei, L., "ImageNet: A Large-Scale Hierarchical Image Database", CVPR, 2009. The origin and scale of the dataset.

AI Now Institute — an institution founded in 2017 at New York University by Kate Crawford and Meredith Whittaker.

Elsewhere

#critical-voices #ai-ethics #invisible-labour #energy-cost #ai-geopolitics