BOOK REVIEW: Luke Fernandez on Karen Hao’s Empire of AI

In her epigraph to The Empire of AI Karen Hao quotes Joseph Weizenbaum who, in the 1960s, invented one of the first chatbots and then became an early critic of AI hype:

It is said that to explain is to explain away. This maxim is nowhere so well fulfilled as in the area of computer programming and artificial intelligence. For in those realms machines are made to behave in wondrous ways, often sufficient to dazzle even the most experienced observer. But once a particular program is unmasked, once its inner workings are explained . . . its magic crumbles away; it stands revealed as a mere collection of procedures. . . .

The quote serves as inspiration for her book. Like Weizenbaum she knows that while AI is magical, that magic veils worrisome power relationships. To explain and unveil those relationships Hao argues that AI is best seen as an empire that is intent on concentrating power and extracting labor and resources from colonized subjects. 

To make her case, Hao focuses on OpenAI and its CEO Sam Altman. In public, Altman is a gracious and charming  pitchman for AI.  He allows that there’s a chance AI might go rogue.  But he assures us that his AI safety team is taking important measures to avoid such an outcome.  Mostly he focuses on AI’s promises and the way it will accelerate intellectual productivity and scientific discovery.  In an early visit to OpenAI’s offices, Hao interviews some of Altman’s co-founders who share much of his vision.  They want to convince her that an AI that is smarter than humans is imminent and that their company intends to share the benefits of those technological advances with everyone.

This dubious vision may rally support for the company. But Hao argues that it glosses over the way power works in empires.  She illustrates this by describing a variety of dramas at OpenAI: the ouster and reinstatement of Altman as CEO, the internal company battles between the doomers who work on “AI safety” (p. 55) and the boomers who seek to commercialize AI, and the departures of Dario Amodei and Elon Musk who have gone on to found their own AI companies (Anthropic and xAI respectively). While these stories of palace intrigue are interesting, Hao is ultimately intent on looking at the empire beyond the castle walls.  She reminds us that OpenAI has imperial ambitions, which it carries out by extracting resources and labor from lands and people far and wide. 

If traditional empires expand spatially, OpenAI expands by scaling up the dataset it uses for training its AI. Instead of using a smaller, more curated, corpus of data, OpenAI decided to use Common Crawl, which contains petabytes of information scraped from the Web.  The choice of quantity over quality means that the inputs are polluted. The data is sometimes copyrighted. Sometimes it contains racist and pedophilic and violent content. Predictably, the tainted inputs yield tainted outputs, which OpenAI then has to clean up using content moderators. The moderators, however, are not working in the same rarified offices as Altman and his close associates. Nor are they salaried workers with health benefits. Instead they are gig workers who often hail from Colombia, Kenya, and other countries from the Global South. As in the book Ghost Work by Mary Gray and Siddharth Suri, Hao visits some of these workers in situ. While they make their own hours, their employment is precarious and pays a pittance. All of this illustrates the important point that, like a traditional empire, the empire of AI extracts labor. And that extraction takes place in locales that extend far from the seat of power. 

The scale of OpenAI’s operations also means that the company needs to house its computers in data centers all over the world. Some of these datacenters are very large.  In Phoenix, for example, three datacenters collectively take up “more than 450 football fields” of land (p. 279). To boot, these datacenters may soon match “the power demands of all of New York City” (p. 280).  Hao cautions that these are estimates – AI companies aren’t transparent about their power needs or their carbon footprint.  But that lack of transparency is itself a worry – to reduce global warming, policy makers and regulators need to know how much power is currently being used and how much will be used in the near future. 

Empires are primarily about labor and resource extraction. But they also have epistemic and ideological dimensions.  In this vein, Hao claims that the victory of connectionism and neural networking (which constitute the primary approaches used by modern AI) have “narrowed the diversity of ideas in AI research” (p. 106).  Since it costs billions to develop a scaled-up LLM (the technology that runs chatbots like ChatGPT, Claude, and Gemini)  the only entities that can afford it are Silicon Valley-backed firms like OpenAI, Google, and Anthropic. As a result, AI research is increasingly difficult to conduct at universities. Hao, rightly, finds this worrisome:

Even as the need for alternatives has grown ever more urgent, the diversity of ideas in AI research has only collapsed further.  Students are dropping out of their PhDs to go straight to industry.  Senior academics are facing a crisis of how to continue pushing the bounds of the field without joining a deep-pocketed company.  More and more researchers have turned their focus not just to deep learning but to large language models exclusively.  The major AI powers are no longer setting the agenda so much as bending an entire discipline to their will. (p. 115)

Here the discipline that is being bent is computer science. But while Hao doesn’t mention it, a strong case can be made that other disciplines are being bent as well. Instructors in writing courses struggle to contend with students who are displacing their own critical thinking and writing skills with chatbot outputs.  Thinking through writing is being displaced by thinking through prompting.  To Hao’s point, the empire extends in many directions – not only is it colonizing labor, land, and energy resources, it’s also colonizing minds. 

At the end of her book, Hao reflects that while extraction may be a critical part of the “formula” (p. 400) for running empires and accumulating power, the most important component may be ideological. She recounts that Sam Altman takes express inspiration from Napoleon and his capacity to rally people around the ideals of liberty, equality, and fraternity even if he wasn’t ultimately faithful to them. In pitching, marketing and hyping OpenAI Altman has been doing something similar.  His promise is that technological progress will make all of us healthy, wealthy and wise. For Altman and many of his co-founders the promise and magic feel real.  His net worth is over two billion dollars. OpenAI has a $300 billion valuation. His co-founder Ilya Sutskever exhorts the denizens of Silicon Valley to “feel the AGI” (p. 120). 

But Hao isn’t seduced by the magic. It fades once Hao steps past the company doorsteps and walks past San Francisco’s homeless population.  It fades even more when Hao recounts how Altman left his sister nearly penniless and resorting to sex work in order to make a living.  And it fades in multiple chapters devoted to the stories of data workers around the globe whose poorly compensated labor is responsible for making supposedly automated systems seem smarter than they actually are. These stories are windows onto the effects of empire on a global scale – the empowerment that is promised to all is concentrated in the few.

Hao’s ethnography and her story isn’t exactly new. Other critics of AI have also raised alarms about the way ghost workers in the Global South labor in “the shadow of AI.” Or how mining, upon which many AI products depend, has adverse environmental effects. But framing it around the thematics of empire is a brilliant choice. She’s a supremely talented narrator and she uses the metaphor with uncommon finesse. 

Too often, AI pitchmen (and their followers) invoke magical thinking. That magic veils the way the AI industry – and particularly OpenAI – is pivotal in deciding how power is distributed in Silicon Valley and around the globe. Hao’s metaphor lifts that veil and allows her readers to see the AI industry and its imperial ambitions and abuses in a broader and more comprehensive light.