The deep dive

Empire of AI

by Karen Hao

18 min read · Original commentary by Lumina, written with AI

An AI answer arrives in seconds. The people and resources behind it are harder to see. Karen Hao follows OpenAI’s rise to ask who directs technological progress. Her challenge is not that useful tools lack value. It is that better products do not, by themselves, give people a fair share of power.

Prefer the quick version? Read the main learnings →

01

Look beyond the chatbot

Empire of AI is reported nonfiction about OpenAI and the wider system that supports generative AI. Its subjects include researchers and executives, but also contract workers, regulators, and people living near data centers. Expect an argument about institutions, not a technical manual or a narrow founder biography.

Hao follows a company founded with a public mission as it becomes more secretive and driven by sales. She does not present this as a simple fall from pure ideals. The early nonprofit already faced questions about openness, competition, and who could speak for humanity. Later financial commitments made those questions harder to contain.

The book asks more than whether OpenAI kept its promises. It asks whether a small group can gather vast resources in humanity’s name without giving other people a real say.

Lumina’s reading is that the key shift comes when AI stops looking like a product made in a lab. Each model belongs to a network of money, work, resources, and rules. Seeing that network lets readers value a tool without placing the same trust in the firm that made it.

A conceptual fork with large server-like blocks on one side and smaller modular blocks on the other, with a hand poised to choose.

Illustrated key idea

Scale is a decision people make

02

Who gets to decide what progress costs?

Karen Hao contends that people chose the dominant path of generative AI; science did not dictate it. Larger models brought real advances. But pursuing them through vast investment, broad data collection, and rapid product launches meant making choices about other people’s lives. Technical success cannot settle those choices on its own.

A useful product can depend on unfair working conditions. A breakthrough can sit behind secrecy that blocks outside scrutiny. Leaders may sincerely expect future benefits while having no clear plan to share them. These concerns do not erase the achievement. They show why achievement cannot answer every question about it.

Hao builds her case by linking choices inside OpenAI to their effects elsewhere. The need for computing power leads toward Microsoft and commercial funding. Model training leads toward collected data and workers who guide answers. Expansion leads toward land, water, electricity, and residents seeking a voice. These links matter more than any single portrait of a powerful leader.

The book also asks how future benefits should count in present decisions. OpenAI’s leaders describe advanced AI as a route to medical and climate gains. Hao asks what supports those hopes and who receives protection while that future remains uncertain. She does not need to prove that the benefits will never arrive. Her challenge concerns the costs people bear while waiting.

Lumina reads this as a question about who may approve a risky trade. Company leaders know their technology. Workers and residents know burdens that a product demo cannot show. Neither kind of knowledge should settle the whole dispute. The book offers no formula for weighing every claim. Instead, it argues that more of the people affected should help set the terms.

03

When bigger models change who holds power

The first major idea is that scale changes institutions as well as software. Bigger models require more data and computing power. Those needs shape funding, partnerships, research access, and which projects survive. A method for raising performance can also make a field depend on a few wealthy firms.

Hao presents scaling laws as patterns found in research results. They linked model size, training data, and computation to measured gains. That gave researchers grounds to expect some improvements from larger systems. It did not prove that expansion would lead to artificial general intelligence, or AGI. OpenAI defined that goal around systems exceeding human performance across most work with economic value.

The difference matters. Better results on a set of tests may justify another experiment. They do not alone establish a route to solving climate change, improving medicine, or sharing wealth. Those are wider claims, with more steps that need evidence. Hao challenges the leap between them rather than denying that models improved.

Microsoft’s 2019 investment makes this link between research and power concrete. Its money and computing systems helped OpenAI pursue expensive work. Microsoft also sought an edge against rivals, whether or not AGI arrived. The partnership brought together different goals. It was not simply a donor funding a public mission.

Lumina sees a cycle here. Expensive research increases the need for large partners. Those partners seek products, customers, or other business gains. Successful products then help fund more research along the same path. Other approaches may lose support without anyone proving them inferior. They may just fit the business less well.

This does not mean smaller systems are always better. Hao recognizes real advances in deep learning, where networks learn patterns from data. The book also describes debate over whether such systems need explicit rules alongside learned patterns. That debate shows that the research path remains contested. It does not establish which rival approach would win with equal funding.

The sharper lesson is about who gets a chance to test an idea. When large firms control scarce computing resources, their choices help set the research agenda. A claim that one path is inevitable deserves special scrutiny when its high costs also shield those firms from rivals.

For readers, the question is therefore twofold: what can scaling achieve, and what forms of dependence does it create? The first is a research question. The second includes business and public policy. Keeping them separate makes it harder to treat a technical gain as permission for every choice surrounding it.

04

The screen hides people who carry the risk

The second idea concerns the human work behind systems that seem to act alone. Users see text or images appear. Hao follows people who sort harmful material, write sample answers, and judge which replies work best. Her point is not just that humans helped build AI. It is that their work can remain unseen while they bear risks others avoid.

One process is reinforcement learning from human feedback, known as RLHF. Contractors provide examples and compare answers to help models respond in useful ways. This makes conversational polish easier to explain without treating it as proof of understanding. Human judgments can improve behavior while leaving false answers unresolved.

The moderation work raises a separate duty of care. Through Sama, OpenAI paid for Kenyan workers to sort texts containing graphic abuse, violence, and self-harm. Some workers said this damaged their mental health and that support fell short. Sama’s response challenged whether workers had alerted it to barriers to counseling. OpenAI, meanwhile, held Sama responsible for weak protections. Hao asks readers to look beyond that division of blame: what did the wider pressure to deliver cheap outsourced work make likely?

That question needs both care and precision. Reports of distress do not settle every claim about what a company knew or provided. Yet a dispute about counseling does not answer who should prevent harm in the first place. If a firm needs people to handle disturbing material, hiring a supplier cannot by itself resolve its duties toward them.

Hao also includes workers who valued the work and wanted better terms, not its removal. This strengthens the argument. Income can be essential even when shifts are unstable or workers have little bargaining power. Gratitude for a job does not require silence about its conditions.

The same problem of visibility extends to physical resources. Data centers need land, electricity, cooling, and mineral-heavy equipment. Near Santiago, residents questioned claims about jobs and local gains. They wanted influence over choices that could affect scarce water. A service that feels remote to its user may be very present in someone else’s neighborhood.

Lumina’s reading is that distance makes responsibility easier to evade. Contracts separate companies from workers; geography separates users from local costs. Neither proves that duties disappear. The book’s task is to make those links open to debate.

The limit matters too. These accounts do not show that every AI tool causes the same harm. Nor do they show that jobs or data centers bring no benefits. They ask for a fuller account, especially when those carrying risk have the least power to change the terms.

A laptop with abstract lines on one side and hands sorting blank cards on the other, connected by a thread to show hidden labor behind code.

Illustrated key idea

AI systems rely on human labor that is easy to miss

05

A public mission needs power behind the promise

Contains plot details — open to read

The third idea separates legal authority from the power to make a decision stick. OpenAI’s nonprofit board was meant to protect its public mission, even against business interests. Hao uses the November 2023 leadership crisis to test that design after staff, investors, and Microsoft had become deeply tied to the company.

For the independent directors, keeping Sam Altman in charge had become a barrier to doing their jobs. They believed his accounts could not be trusted and that he had blocked effective oversight. Their decision rested on a collection of incidents and reports, not one startling technical discovery. Concerns raised by Mira Murati and Ilya Sutskever added weight to their own experiences. These remain the directors’ judgments, not a public verdict establishing every allegation.

They removed Altman but gave little detail about their reasons. Secrecy had protected their talks and people who supplied information. Once the decision became public, that same secrecy weakened their ability to explain it. Employees feared for their work, the company’s future, and valuable equity. Investors supported Altman. Microsoft offered to hire him and departing staff, giving opponents of the board somewhere else to go.

More than 700 employees signed a letter seeking Altman’s return. Murati and later Sutskever backed that demand. A settlement brought him back under a changed board. Hao treats the reversal as evidence that nonprofit control could not withstand the company’s business ties and internal pressures.

Lumina sees two distinct failures to examine. First, can directors get sound information and act on it? Second, can they sustain an action once powerful people resist? A board may have the right to remove a leader but be without staff trust, funding options, or independent access to computing systems.

This does not require defending every choice the directors made. Their sparse explanation helped fuel confusion. Nor does the revolt show that staff cared only about money. Hao describes loyalty, fear of collapse, belief in the work, and financial stakes. A single-motive story would miss what made the crisis hard to control.

The later review offers no simple public resolution. Its report stayed private, and its focus did not openly settle the wider dispute over leadership and accountability. Hao also excludes secret reasoning research as the cause of the firing. That prevents a neat but unsupported story about a hidden breakthrough.

The wider lesson is not merely to choose better leaders. Oversight needs reliable information, independent resources, and rules that survive conflict. Even a board victory would leave Hao’s larger question open: why should choices affecting so many people turn on which small group inside a company wins?

06

Hao connects the boardroom to distant lives

Contains plot details — open to read

Hao’s main storytelling move is to widen the frame. The opening leadership fight creates urgency, but she does not stay inside it. She follows the money, work, and resources that support the company. Lumina reads this movement as part of her argument: an internal choice becomes a public concern when its wider effects come into view.

The 2019 office visit offers one clear anchor. Brockman and Sutskever defend advanced AI through possible gains in medicine and climate work. Hao asks why those goals demand a general technology that does not yet exist. Useful machine-learning tools already face barriers that involve policy, adoption, and expert cooperation. Her questions separate belief in a goal from evidence that a chosen route is needed.

The labor accounts work differently. Named people give a human scale to terms such as data labeling and moderation. Their experiences disrupt the idea that useful model behavior comes from computation alone. Company replies keep some details in dispute. The argument need not turn every participant’s account into settled fact to remain serious.

Hao also explains why secrecy can weaken scientific claims. A test is less revealing when a system has already encountered its answers. Training records would help outside researchers assess that possibility. Without those records, a strong score can support several explanations, and the company’s preferred one is harder to check. This limits what a result proves; it does not make every score worthless.

The empire comparison joins these strands. Hao uses it to examine how firms gather resources and rewards while shifting costs outward. She does not equate those practices with colonialism’s overt violence. Readers should preserve that limit rather than letting the title stand in for evidence.

Lumina’s caution is that a broad frame can gather too much momentum. Readers may start seeing every dispute as proof of the same pattern. A stronger reading asks where each link rests: on a documented event, a participant’s account, or Hao’s interpretation. The book’s full reporting matters because those differences deserve more attention, not less.

07

Can wider control keep the benefits of ambition?

The strongest fair objection, in Lumina’s view, is that Hao’s criticism is more developed than her replacement. Major research can require large sums, skilled teams, and close coordination. Community control sounds appealing, but that does not show how it would manage every demanding project. Exposing the costs of concentrated power is not the same as proving another system can retain its benefits.

Te Hiku Media offers an alternative within a clear task. Its Māori speech-recognition work rests on community consent, control over data, and a language-preservation goal. It used limited computing resources. The example shows that useful AI need not start with broad extraction. It does not prove that this model can match the range of a frontier general-purpose system.

The strongest reply is that matching that range may be the wrong test. Hao asks whether maximum scale should define success for every use of AI. A tool built around a community’s need can matter without replacing OpenAI. On this reading, we should ask whether a system serves its purpose on fair terms, not only whether it rivals the largest firms.

A second tension concerns future danger. Hao argues that extreme scenarios can draw attention away from current labor, privacy, bias, and environmental problems. A fair objection is that present harms do not make future threats trivial. Waiting for conclusive evidence of a grave risk could itself be reckless.

Lumina’s reply is that these concerns need not compete. Fear of future systems could support stronger oversight today. The problem comes when it gives more power to a narrow group while leaving present duties weak. Hao’s case is strongest as a challenge to selective accountability, not as proof that future danger is imaginary.

Disclosure also has limits. People need time, skills, money, and enforceable rights to act on information. Hao’s calls for independent research, labor rules, and public education address part of that problem. What remains uncertain is how those supports gain enough force against firms able to move investment elsewhere. The alternatives are real openings, not a demonstrated final answer.

08

Ask about purpose, voice, and enforceable rights

Scale, hidden work, and weak oversight connect through a practical question: can affected people change a decision? Seeing the costs is only a start. A promise to serve the public is not enough either. Lumina suggests three linked tests drawn from the book’s concerns: purpose, voice, and enforceable rights.

Purpose asks what problem a system should solve. Hao challenges the jump from a real social need to a claim that AGI is required. A clear goal allows comparison with simpler or less costly tools. It also helps reveal when growing a company has become a separate aim from meeting the need.

Voice asks whose knowledge shapes the plan. Workers know the demands of disturbing content and uncertain shifts. Residents know local water pressures and whether promised jobs last. Researchers know gaps in testing. The book’s community examples suggest that consultation should help set terms, not merely explain choices already made.

Rights ask what happens when promises clash with pressure to grow. Can workers question their treatment without losing income? Can outside researchers inspect enough evidence to test a claim? Can an oversight body stop a release without being bypassed? These questions link the supply chain to the boardroom.

For a brief imagined example created by Lumina, suppose a community is offered an AI service along with new infrastructure. A good demo might show that the service works. It cannot show that water demands are fair, workers are protected, or residents have a right to appeal. Those claims need their own evidence and agreements.

This approach does not sort all technology into good and bad. A useful system may need stronger rules rather than rejection. A worthy goal may still be pursued on unfair terms. Keeping those judgments separate preserves the book’s central opening: people can debate the direction of AI instead of treating its current business model as fate.

A cloud icon above a simplified building connected by pipes to a water container and an electric pylon shape.

Illustrated key idea

Cloud AI depends on local resources

09

Keep disagreement focused on claims we can test

A good conversation separates what a tool can do from who has the right to set its terms. Someone may value a chatbot while objecting to how it was built. Someone else may support ambitious research while seeking stronger labor rules. Hao’s argument leaves room for both because it targets the direction of AI, not every use of it.

When someone calls progress inevitable, ask what they mean. Is it a technical gain, a funding model, or control by a few firms? Evidence for one does not establish the others. When someone invokes safety, ask whose safety counts and what action follows. When someone calls for public control, ask which choices people could change.

Lumina also suggests naming the kind of claim at issue. Hao reports events, presents conflicting accounts, and offers an explanation that connects them. Agreeing with her broad case should not erase those differences. Disagreeing should identify a weak link, not dismiss every concern because some products help people.

This deep dive prepares discussion of the argument; it does not replace the experience of reading the reporting. A strong follow-up stays concrete: what change would give people real authority, and what evidence would show that it worked? That keeps benefits and duties in the same conversation.

What to remember

  1. Technical success does not settle who should hold power.

    Scaling brought gains but also raised the need for money and computing systems. Hao asks readers to examine the ties that followed. Better performance supports a claim about a model. It does not automatically justify secrecy, concentrated control, or promises that everyone will benefit.

  2. A contract cannot answer every question about responsibility.

    Human feedback, moderation, and data centers connect AI services to people with little bargaining power. Lumina’s synthesis is that hiring a supplier does not settle who owes protection. Useful jobs and products can still rest on terms that workers and residents have reason to challenge.

  3. Public benefit needs people who can shape the terms.

    Hao’s alternatives include community control of data, independent research, disclosure, and worker rights. Their shared aim is not simply smaller models. It is to let affected people influence goals, costs, and remedies instead of relying on a company’s account of what serves them.

Bring it into conversation

Talk about the book

What would justify calling large-model development necessary rather than promising?

Hao separates measured gains from hopes about general intelligence. One can support scaling while still asking for fair tests of other approaches and a fuller account of costs. The evidence should fit the size of the claim, not just the appeal of the goal.

Take it further: Which result would change your view, and which choices about funding or control would remain open?

Who should decide when users gain but workers or local residents bear costs?

The book rejects overall usefulness as the end of the debate. Yet giving every group equal say may not resolve conflicts. A serious answer needs to name rights, forms of representation, and routes of appeal rather than invoke participation as a cure for everything.

Take it further: Which choices require consent, which require negotiation, and who should enforce that difference?

Must community-led AI rival general-purpose models to count as a strong alternative?

Te Hiku Media offers another measure: serving a clear need under community rules. It does not prove every large project can use the same approach. It does challenge the idea that broader capability is always more valuable than control over purpose and use.

Take it further: What goals become easier to notice when size and market reach stop being the main tests?

Keep in perspective

This analysis follows Hao’s account; it does not independently verify disputed allegations, private reviews, or technical findings. The available material provides no wider author biography or outside review debate. Community projects show other paths are possible, not that they can replace every frontier system. Later industry changes fall outside the account examined here.

Who's talking about Empire of AI

Public statements we checked against the original source

Portraits from Wikimedia Commons · photo credits and licences