Empire of AI; Review

What It's About

Karen Hao is a journalist who has covered OpenAI since 2019. Her book says the company runs like an old empire. It takes labor and resources from poor countries and keeps the money. Then it calls the whole thing a mission to help humanity. She spends four hundred pages backing that up with names, numbers, and interviews.

Hao lays the pattern out early. An empire claims a territory and calls the taking a mission to civilize it. It pulls cheap labor and raw material off the land, and the wealth goes home to a small group that calls itself generous. Then she walks through OpenAI's history and checks the company against that pattern, step by step. The pattern holds the whole way.

The mission is artificial general intelligence (AGI), a computer that could think as well as a person at any task. AGI stays a few years away no matter what year it is. Inside the company, one camp believes AGI will save the world. Another camp believes it could end it. Both camps agree that OpenAI alone should build it, and that the normal rules should not apply along the way. The company started as a nonprofit with a safety board sitting above the CEO. Then Microsoft put in thirteen billion dollars. In Nov 23 the board fired Sam Altman. Its own public statement said he had not been "consistently candid" with them. He was back inside a week, and the board members who fired him were gone instead. That is what thirteen billion dollars buys.

What I Think

One chapter is about a man named Mophat Okinyi. He lives in Nairobi, Kenya. A staffing company called Sama hired him to work for OpenAI. His job was reading the worst text the model could produce, hour after hour, so ChatGPT would learn to refuse it. The pay was under two dollars an hour. Billy Perrigo, a reporter at TIME magazine, published that pay number first. Hao goes further. She reports what the work did to Okinyi, and he did not come out of that contract the same man who went in. I set the book down and sat with that for a while.

I spent years around classified systems where the data could not leave the building. That work teaches you a habit. You stop asking what the machine can do. You start asking who is doing the work. Hao went and asked. She flew to Kenya, to Venezuela, and to the data center towns of central Chile. She sat with the people who did the work and she wrote down what they told her. Most coverage of this industry never leaves the press release. This book was reported the old way, in person, and it reads that way. That is why I trust it.

What It Gets Right

Here is how the product got made. People posted answers on the internet for free, millions of them, for years. They were just helping each other out. OpenAI copied all of it. Then it hired workers in Kenya and Venezuela to clean up the raw output for a dollar or two an hour. Then it sold the result and called it machine intelligence. The company is worth almost a trillion dollars now. The computer did not create that knowledge. People did, and the company packaged it.

Think about what a subscriber is actually paying for. The answers came from forum posts and help threads that people wrote for free, to help a stranger. The cleanup came from workers earning a dollar or two an hour. The monthly fee pays for the package. Almost none of it reaches the people who made the thing work.

Hao tracked the workers down. Okinyi earned less in an hour than one California subscriber pays in a month. In Venezuela the economy collapsed, and people with college degrees took labeling jobs because nothing else existed. Then the work disappeared. Hao reports that Scale AI, a company that runs data labeling platforms, shut down its Remotasks platform in Kenya in Mar 24. Thousands of workers lost their income overnight with no warning at all.

There is a water problem too. Data centers use huge amounts of water for cooling, and new ones are going up in central Chile and Uruguay. Chile has been in a drought for over fifteen years. The towns hand over land and tax breaks. The water goes to the servers, and the profit goes to another country.

She is fair to the believers, and the book is stronger for it. She lets the AGI faithful talk at length, and the faith comes through sincere. These are engineers who honestly think they are saving the world. She never sneers at them. Some employees told her the company was building the most important technology in history. Others told her it might be the most dangerous. She quotes both camps at length and lets the reader hear how certain everyone is. Then she just keeps reporting until you see the problem for yourself. If you believe you are saving the world, you never have to explain the wage in Kenya or the water bill in Chile.

What It Gets Wrong

Two problems, and both are gaps.

The first is math. Her biggest number for data center water use in Chile looks too high. A writer named Andy Masley checked it in print, and the number did not hold up. The worry underneath it is real. One facility asked for more water than a nearby town uses, in the middle of a long drought. But a case this strong needs airtight numbers. The companies employ people whose whole job is finding the weak spot in a critic's math. One soft number gives the industry a comeback it did not earn, and the other four hundred pages get ignored in the noise.

The second is where her fix stops. She points to Te Hiku Media, a Māori community group in New Zealand. They trained language tools on recordings of their own elders speaking their own language. They wrote the rules for who can use the results, and the community owns all of it. That is the right model, and it is a real one. She leaves it at community projects. The same idea works one level up, for the small business and the trade shop that could run their own tools on their own terms. That chapter never comes. In fairness, it was never hers to write.

What I'd Add

A small model trained on one trade, run by somebody who knows that trade, beats a giant rented one run by somebody who doesn't. I have worked systems where the data could not leave the building, and the lesson from that work is plain. The model is a tool, and the operator decides what the tool is worth. The industry hides the operator, because a hidden operator is what sells the product.

Here is what the other way looks like. A fire alarm company could train a small model on its own service records and its own panel documentation. That model would answer questions about the exact systems the company maintains, and nothing else. The data never leaves the shop, and the shop owns the result. None of that requires a trillion dollar company. It requires a used graphics card and somebody who knows the work.

The companies have started admitting it themselves. In spring 25 the CEO of Anthropic said AI would write ninety percent of code within six months. IT Pro, a technology news site, checked that claim in Sep 25, and the real number came in near thirty percent. The CEO of IBM told TechCrunch, another technology news site, that twenty to thirty was closer to the truth. Fortune reported this past May that Sam Altman said he was "delighted to be wrong" about the job losses he had spent years predicting. The entry level jobs he expected to vanish were still there.

Now look at the calendar. Fortune reported that Anthropic filed to sell shares to the public on 1 Jun 26. TechCrunch reported that OpenAI filed on 8 Jun 26. OpenAI's own forecast, reported in Jan 26, shows a fourteen billion dollar loss this year with no profit expected before 2029. The doom talk got walked back the same season the offering paperwork went in. Both companies are hiring programmers as fast as they can.

Verdict

This is the most honest account I have read of what this industry is and what it costs. Hao interviewed the executives who run OpenAI, and she interviewed the workers who clean up after its model. Almost nobody does both. I finished the book trusting the reporter more than her subject, and I think that was her goal.

The work was human the whole way through, and the people who did it belong in the story. Nobody has written the version of this story for small shops running the technology on their own terms.