Kate Crawford is a researcher who has spent a long time studying artificial intelligence as a social and material phenomenon rather than a technical one, and this book is the fullest statement of that position. The organising idea is in the title. It is an atlas, a set of maps of the territory that AI actually occupies, and that territory turns out to be mines, factories, logistics networks, call centres, classification schemes and government agencies rather than the abstraction most people have in mind. The opening move is the most effective in the book.
She goes to a lithium mine, and describes it. Not as a metaphor and not as a statistic, as a place with an environmental cost and a labour situation and a supply chain running out of it. The point is simple and lands hard. Every model runs on hardware, the hardware is made of minerals, the minerals come out of the ground somewhere, and the abstraction we work in is sitting on top of an extraction industry that nobody in the field thinks about.
The chapter on classification is the one I would make people read. It examines how training sets get built, what happens when you decide on a set of categories and label the world according to them, and what those categories carry. The analysis of the large image datasets that trained a generation of computer vision systems is careful and damning. Categories that made no sense, categories that were straightforwardly offensive, hierarchies inherited from linguistic databases never designed for the purpose, and all of it presented downstream as ground truth.
Anyone who has ever used a benchmark dataset without asking where the labels came from should read this and then go and look. The labour reporting is strong. Crawford follows the human work that sits inside supposedly automated systems, from warehouse operations to data annotation to content moderation, and shows how consistently the industry describes as automatic something that has people inside it doing the parts the system cannot. This is the same territory more recent books have covered with more current examples, and Crawford was making the argument earlier and connects it to the extraction material more tightly.
The chapters on affect recognition and on the state's use of these systems are solid. The affect material is the best short critique available of the claim that emotion can be read from a face, tracing the shaky psychological foundations that the technology inherited and never examined. Now the criticisms, and the first is about the writing. This is academic prose.
Long sentences, theoretical vocabulary, a register that assumes familiarity with a particular scholarly tradition. The arguments underneath are clear and forceful and the delivery puts distance between them and the general reader. There is a version of this book that reaches ten times the audience by being written more plainly, and this is not that version. Engineers in particular tend to bounce off the style before reaching the substance, which is a shame because they are the people whose work it describes.
It is dated in an important way. Published in 2021, before generative models became the centre of the field. Everything here is about classification, prediction and recognition systems, and the extraction and labour arguments transfer to the current generation, and the specific analysis does not. A reader wanting to think about large language models will have to do that transfer unaided, and the book's examples will feel like they are from a previous era, because they are.
The technical material is the weakest part. Where Crawford describes how systems work, the account is thin and occasionally loose in ways that a practitioner will notice. This does not undermine the central arguments, which are about political economy rather than architecture, and it does cost credibility with exactly the technical audience the book most needs to reach. A few pages of tighter technical writing would have closed that gap.
It is not actionable. The book diagnoses and does not prescribe. There is no method here for auditing your data, no framework for making a better decision, nothing you can apply. That is a legitimate choice for a work of critical scholarship and it means a practitioner finishes it convinced there is a problem and with no idea what to do differently.
The argument is one sided and does not really pretend otherwise. Counterarguments are mentioned and not seriously engaged. The strongest case for these systems, that they deliver real benefits in medicine, accessibility, translation and science, is largely absent, and a reader who wants a balanced assessment will need to supply the other half themselves. Where the book is right, it is right, and it is arguing a position rather than weighing one.
My three point six is for a genuinely valuable reframing that will change how you look at a dataset and a supply chain, delivered in a style that limits its reach, dated by a field that moved underneath it, and weakened by technical looseness and a reluctance to engage with the other side. Read the classification chapter and the labour chapters. Those are the parts that will stay with you.