Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Book Review)

By jgn on Saturday, June 28, 2025 in Reading, Reviews, and Technology

This book shares the story of the rise of OpenAI, and is thoroughly convincing on a number of fronts: (1) OpenAI betrayed its original organization as a mission-driven non-profit wanting to create "safe" AI, turning part of itself into a for-profit entity; (2) OpenAI reduced its focus on safe AI; (3) OpenAI stopped releasing its work in public, taking its models private. In those respects, OpenAI is a failure, at least with respect to its founding vision.

Then there is the story around Sam Altman, OpenAI's CEO, who seems to have remarkable skill telling people what they want to hear while doing what he wants. Hao makes a pretty good case that Altman is an incorrigible liar.

The book does a good job narrating the ups-and-downs of OpenAI's governance, largely through the activities of Altman. Particularly toward the end of the book, OpenAI's leadership was a mess, with arguments between the board and the CEO, conflicts within the board, departures of leadership, and especially departures from the safety group. There's some good muckraking here, and the story is compelling. Sam Altman is in the title, but he's not in every chapter -- I think that's a problem for the book, which doesn't quite know what it's doing.

The book's overarching argument, such as it is, is pretty dubious. I almost scored it a 3 on this basis. The problem is that there is an argument in the book that OpenAI acts as an "empire" and analogies are drawn between the empires and colonial ambitions of 19th and 20th centuries. There's a good chapter on the way AI exploits low-wage temp workers in Africa and South America (chapter 9, "Disaster Capitalism"). But it isn't as strongly tied to OpenAI (and no mention of Altman) as it might be. You could easily argue that the problem is late capitalism rather than AI. You get a similar implication about very large scale data centers (chapter 12, "Plundered Earth"): but again, is it really AI here, or really OpenAI in particular (again, Altman doesn't appear)? To make the argument that AI per se is the root cause here would be interesting, but it isn't made in any strong sociological sense. The corporate behavior is terrible, but this seems like a demand-side problem about the rapaciousness of the market. There's a nice epilogue that talks about a case that seems to "get it right" with an AI project that is very public, respects resources, and asks permission of people to include their data for an effort to revitalize the te reo Māori, the language of the Māori people. Hao does a nice job differentiating between the "do it all" models of OpenAI and Anthropic in favor of smaller more specialized models: The infrastructure for the te reo project requires only two GPUs, whereas ChatGPT, OpenAI's key product, requires thousands. And it used a model based on DeepSeek (the book seems to have gone to press in January 2025 so missed the big blowup about DeepSeek's chip efficiency, a direct threat to the economics of OpenAI).

In short: A lot of great stuff, but the attempt to paint OpenAI as structurally a cause of misery around people and resources didn't convince me. It's terrible but making the argument requires a lot of steps. Another book probably would make the case (perhaps it will be The Age of Extraction: How Tech Platforms Conquered the Economy and Threaten Our Future Prosperity).

Thanks to the Claude AI for tweaking some spelling and grammar.

Originally reviewed on Goodreads on 28 June 2025.