Karen Hao covered artificial intelligence at MIT Technology Review, embedded with OpenAI for a profile in 2020, and has spent the years since reporting on the industry from the outside after that relationship soured. That history matters, because it produced a book that is neither an authorised account nor an outsider's guess. She has spoken to a very large number of people, many of them on the record, and the reporting is the reason to read this even if you disagree with every conclusion she draws from it. The central argument is in the title.
Hao proposes that the way to understand the current AI industry is as an imperial project, extracting resources and labour from places with less power to build systems that concentrate wealth and capability somewhere else. Compute and capital in a handful of companies, training data taken without consent from everyone who has ever published anything, content moderation and data labelling done for very little money by workers in Kenya and Venezuela and the Philippines, energy and water drawn from communities that get datacentre construction and not much else. She argues this is not a set of unfortunate side effects but the structure of the thing. The OpenAI material is the spine of the book and the part with the widest appeal.
The nonprofit founding and the reasoning behind it, the drift towards a capped profit structure, the Microsoft relationship, the tension between the safety mission and the commercial pressure, and the board crisis of November 2023 told in more detail than anywhere else I have read. Hao had sources on multiple sides of that and the account is specific about who believed what and when. Anyone who followed that week through social media and news alerts will find the full version genuinely clarifying. The labour reporting is the part I would press on people who think this is a book about tech gossip.
Reinforcement learning from human feedback requires humans giving feedback, and content classifiers require humans labelling the worst material on the internet so a model can learn to refuse it. Hao follows that work to the people doing it, what they were paid, what they saw, and what happened to them afterwards. This is not an abstract ethics discussion. It is a description of a production process, and it is a production process that every model any of us use depends on.
Most engineers have no idea it exists in the form it does. The environmental chapters are similarly concrete. Water consumption for cooling, grid pressure from datacentre construction, the local politics of where these facilities get built and who gets a say. She goes to places and talks to residents, which is more than the numbers in most coverage will give you.
Now the honest criticisms. The empire frame is a strong analytical device and Hao commits to it fully, which means it sometimes gets applied to situations that fit awkwardly. Colonialism is a specific historical thing with specific mechanisms, and stretching it to cover every power asymmetry in a technology industry costs some precision. The comparison is illuminating often enough to earn its place in the book, and it is doing more work in some chapters than the evidence supports.
This is also not a book about the technology. You will not learn how a transformer works, why scaling produced the capabilities it did, or anything about the research decisions beyond who made them and what the consequences were. That is a legitimate scope choice for a book about an industry rather than a field, and it does mean a reader hoping to understand the systems themselves needs something else entirely. Hao writes from a position and does not hide it, which I prefer to false neutrality, and it means you should read critically.
Where a decision could be read as either cynical or reasonable, the book tends to reach for cynical. Some of those readings are well supported. Some rest on the interpretation of a source who had a reason to interpret it that way. A careful reader can usually tell which is which, because the reporting is laid out well enough to disagree with.
The pace of the field is the unavoidable problem. This came out in 2025 and the corporate landscape has moved since. Structures have changed, people have left, new labs have raised extraordinary amounts of money. The specifics will keep dating and the structural observations will not, which is the usual bargain with a book about a live subject.
Why it belongs on a site about learning. Almost everyone in this field learns the technology and never learns the industry. You can be an excellent engineer with no idea where your training data came from, who cleaned it, what it cost to run the cluster, or what the governance structure of the company whose API you depend on actually looks like. That is a real gap and it affects decisions, from which providers you build on to what you are willing to put your name to.
This book fills a large part of it. My four point four is for outstanding reporting on the parts of this industry that are hardest to see and most rarely covered, marked down for a framing device that gets pushed harder than it needs to be and for a subject that will not hold still. Read it alongside something technical rather than instead of one.