Atlas of AI

Crawford's structure is a journey through the supply chain: lithium mines in Nevada, data centres and their water consumption, the click-work that labels training data, the datasets assembled from photographs taken without consent, and the state power that the whole thing serves.

Her claim is that treating AI as immaterial software obscures an industry with a physical footprint and a labour force, both of which are distributed according to existing inequalities.

The book's method is to follow the material: where the lithium comes from, who labels the data, who pays for the water.

Devmark’s reading — not a quotation from the book

The section on training datasets is the most concrete — how image corpora were assembled from mugshots and scraped photographs, and what the classification schemes assumed about the people in them.

The labour chapter on crowdwork is similarly specific: the human classification that makes automated systems possible, priced and distributed to be invisible.

Crawford is a researcher at USC and Microsoft Research and has been inside these systems, which gives the critique specificity that outside accounts lack.

It is a critical book and makes no attempt to be balanced about benefits. Read it as the material counterweight to technical accounts rather than as a survey of the field.

It is a counterweight, not a survey, and does not pretend otherwise.

Devmark’s reading — not a quotation from the book

Crawford's deeper objection is to the word intelligence itself. She argues that describing statistical pattern-matching as cognition imports a set of assumptions about minds that the systems do not earn, and that the metaphor then shapes regulation written by people who accepted it.

That is the thread connecting the mines to the datasets: each chapter shows a material dependency that the language of artificial intelligence conceals.

The opening chapter is set at a lithium operation and is doing argumentative work: before any discussion of models, Crawford establishes that these systems begin in extraction — minerals, water, land, and the geopolitics attached to each. Her point is that an industry describing itself as immaterial has an unusually heavy physical footprint that its own vocabulary is designed to obscure.

The labour chapter documents the people whose work is presented as machine output: warehouse workers managed to the second, crowdworkers labelling training data for fractions of a cent, and systems marketed as autonomous that are substantially human behind the interface. She links the measurement regimes directly to early twentieth-century scientific management, arguing the technology is new and the logic is not.

The autonomy is frequently a person, paid by the task, kept out of the photograph.

Devmark’s reading — not a quotation from the book

The most concrete material concerns provenance: research corpora assembled from mugshot archives, from photographs scraped without consent, and from images of people who never knew they were in a dataset. Crawford's argument is that a technical field treated this material as freely available because it was accessible, and that the resulting norms are now embedded in systems used on the people depicted.

The chapter on taxonomy examines the category structures inside benchmark datasets and finds a large number of judgements about people — occupational, moral, and worse — presented as objective labels. Her claim is that the act of sorting the world into classes is where politics enters, long before any model is trained, and that benchmarks inherited those categories uncritically.

One of the strongest chapters attacks affect recognition. The commercial systems rest on a mid-century theory of a small set of universal emotions with fixed facial signatures, and that theory has been substantially contested within psychology on the grounds that expressions vary by culture, context and individual. Products are being sold on a scientific claim that is disputed at its foundation.

The final movement follows these systems into intelligence gathering, policing and immigration enforcement, arguing that the capabilities and the institutions developed together rather than one being applied to the other afterwards. It is the section where Crawford's position as a researcher who has worked inside major labs gives the account unusual weight.

It is a critique from the humanities and social sciences rather than a technical book, and readers wanting engineering detail will not find it. What it supplies is a durable question to put to any AI claim: what is being extracted, from whom, who labelled it, who is inside the automation, and who carries the cost. That survives whatever the models do next.