Inside OpenAI’s Empire: Sam Altman’s Power Struggles, the Jobs AI Is Replacing, and the Hidden Cost of the Race

In November 2023, OpenAI's board fired Sam Altman, the man who had transformed a nonprofit research laboratory into one of the world's most influential technology companies. Five days later, he was back.

Investigative journalist Karen Hao argues that the episode revealed something larger than a failed boardroom rebellion. It exposed how money, personal influence, scientific ambition, and competing visions of artificial intelligence had become entangled inside the company behind ChatGPT.

In a wide-ranging conversation with Steven Bartlett on The Diary of a CEO, Hao examined OpenAI's internal conflicts, Elon Musk's departure, the treatment of critical researchers, the workers training AI systems, and the communities supplying the electricity and water needed to operate them.

Her book, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI, draws on more than 300 interviews with over 250 people, including more than 90 current or former OpenAI employees and executives.

The central question is not whether artificial intelligence is useful. It is who controls its development, who receives its economic benefits, and who bears its costs.

1. How Sam Altman Lost Control of OpenAI — and Won It Back

The warnings that reached OpenAI's board

By 2023, OpenAI's chief scientist, Ilya Sutskever, had become increasingly concerned about Sam Altman's leadership.

Sutskever had two overriding ambitions: developing artificial general intelligence, or AGI, and ensuring that such technology would be safe. According to Hao's reporting, he came to believe that Altman's management was undermining both.

He reportedly saw a company becoming increasingly chaotic. Research teams were competing rather than cooperating, executives were receiving conflicting information, and internal trust was deteriorating.

Sutskever eventually approached Helen Toner, an independent member of OpenAI's board. Hao says he was initially cautious about sharing his concerns because he feared the consequences if Altman discovered what he was saying.

Mira Murati, OpenAI's chief technology officer, also raised concerns with directors. According to Hao, the discussions included emails, Slack messages, and examples of allegedly contradictory communications and internal rivalries.

These were significant complaints because OpenAI was not governed like an ordinary technology startup.

Founded as a nonprofit in 2015, OpenAI had established a commercial subsidiary in 2019 while retaining nonprofit oversight. Independent board members were meant to protect its public-interest mission rather than maximize returns for investors.

Directors therefore had a formal responsibility to challenge management when they believed the organization's mission was being compromised.

ChatGPT's success created a company OpenAI was not prepared to manage

ChatGPT launched in November 2022 as a research preview based on GPT-3.5.

According to Hao, OpenAI had expected to gather feedback before eventually introducing a more capable chatbot based on GPT-4. Instead, ChatGPT became an immediate mass-market phenomenon.

The company struggled with server capacity, outages, rapid recruitment, and the demands of operating a service used by millions of people.

Hao describes aggressive hiring followed by abrupt departures. In some cases, employees reportedly learned that colleagues had left because their Slack accounts suddenly disappeared.

The management concerns emerged against this backdrop.

OpenAI was simultaneously attempting to advance AI research, operate a rapidly growing consumer platform, satisfy investors, and maintain its safety commitments.

Sutskever and Murati believed the resulting organizational instability was becoming dangerous.

The Startup Fund raised another governance question

One issue attracted particular attention from OpenAI's independent directors: the legal structure of the OpenAI Startup Fund.

According to Hao, board member Adam D'Angelo began asking questions after hearing concerns about the fund at a San Francisco gathering.

Although the investment operation appeared to be a corporate venture fund associated with OpenAI, documents showed that it was legally controlled through Altman personally.

The distinction mattered. Directors overseeing a mission-driven organization needed to understand which assets, investment operations, and financial arrangements were legally controlled by the organization and which were not.

Independent reporting subsequently established that the arrangement was unusual.

In February 2024, Axios reported that the fund was legally controlled by Altman, rather than OpenAI itself.

The fund had received approximately $175 million in commitments from outside investors, including Microsoft.

OpenAI maintained that Altman had no personal financial interest in the fund and that the arrangement had been intended to be temporary.

In April 2024, control was transferred to fund partner Ian Hathaway.

The legal structure was real, but it would be incorrect to conclude from that structure alone that Altman personally owned the underlying investments for his financial benefit.

For the directors, Hao says, the issue contributed to a wider loss of confidence in the accuracy and completeness of information supplied to the board.

Why directors acted without warning

By November 2023, independent directors had concluded that Altman should be removed.

Hao says the decision was influenced partly by their fear of his extraordinary ability to persuade people.

They believed that informing him too early would allow him to mobilize employees, investors, and influential supporters against the decision.

On November 17, 2023, OpenAI announced his removal.

Its official statement said the board had concluded that Altman had not been consistently candid in his communications with directors.

The announcement gave little detail about the underlying disputes.

Microsoft, OpenAI's most important commercial partner, received almost no advance warning. Employees and investors were similarly unprepared.

The secrecy that directors believed would protect their decision helped undermine it.

Microsoft, investors, employees, and other stakeholders rapidly mobilized in support of Altman's return. Within days, the board's position became politically and operationally unsustainable.

Altman returned as chief executive on November 22, 2023.

The people who had challenged him paid a significant professional price. Sutskever did not return to his previous position and subsequently left the company. Murati also departed.

Hao's book recounts Sutskever privately expressing a profound loss of confidence in Altman. Bartlett repeated the reported remark during their interview:

"I don't think Sam is the guy who should have the finger on the button for AGI."

The statement reflects Hao's account of Sutskever's concerns, not an independently released recording of his private conversation.

What OpenAI's independent review established

The boardroom dispute did not end with Altman's reinstatement.

In March 2024, OpenAI released findings from an independent review conducted by the law firm WilmerHale.

The firm reviewed more than 30,000 documents and interviewed numerous witnesses.

Its published findings confirmed that a breakdown of trust between Altman and the previous board had triggered the dismissal.

However, the review also concluded that the dismissal had not resulted from concerns about product safety, security, the pace of AI development, company finances, or representations to investors and customers.

It found that directors had acted within their authority but had proceeded without sufficient inquiry, advance notice to stakeholders, or an opportunity for Altman to respond.

The reviewing board expressed confidence in Altman and Greg Brockman and introduced governance changes, including stronger conflict-of-interest procedures and a whistleblower mechanism.

These findings complicate any simple account of the controversy.

Hao's reporting emphasizes the experiences and concerns of executives who challenged Altman. OpenAI's review acknowledged the breakdown in trust but did not conclude that his conduct required removal.

The documented result was that Altman returned with substantially stronger leadership authority, while the directors who had attempted to remove him lost control of the organization.

2. Elon Musk, Sam Altman, and OpenAI's Original Power Struggle

OpenAI began with a warning about dangerous AI

OpenAI's founding in 2015 involved an unusual combination of technological optimism and fear.

Sam Altman and Elon Musk both publicly warned that sufficiently advanced AI could pose an existential threat to humanity.

Musk famously compared developing advanced AI to summoning a demon.

In a 2015 blog post, Altman described superhuman machine intelligence as potentially the greatest threat to humanity's continued existence.

Hao argues that Altman's language closely mirrored Musk's public warnings at a time when he was attempting to recruit Musk as a co-founder.

Before that period, she says, Altman had emphasized threats such as engineered viruses.

Her interpretation is that Altman adapted his language to persuade Musk that the two men shared an understanding of AI's dangers.

That interpretation should not be mistaken for proof that Altman's concerns were insincere. A person's language can serve a strategic purpose while reflecting a genuine belief.

Musk joined the effort and provided substantial financial support.

The dispute over who would run OpenAI's commercial operation

The founding nonprofit soon confronted an expensive reality: advanced AI research required increasingly large quantities of computing power.

By 2017, OpenAI's leaders were discussing a commercial structure capable of attracting billions of dollars.

According to Hao, Sutskever and Brockman initially supported Musk becoming chief executive of the proposed for-profit entity.

Altman privately challenged that idea.

Hao reports that he approached Brockman, a longtime associate, and questioned whether Musk's fame, outside business commitments, public pressures, and unpredictability made him a suitable person to control potentially powerful AI technology.

Brockman reconsidered and persuaded Sutskever to change his position.

Musk was no longer the preferred candidate.

Hao presents this reversal as a critical event in the collapse of the founders' relationship.

Documents subsequently published by OpenAI support the broader conclusion that the dispute concerned both financing and control, although they do not independently establish every private conversation described by Hao.

In its December 2024 account, accompanied by historical emails, OpenAI said Musk had proposed a commercial arrangement in 2017 that would give him majority ownership, extensive control, and the chief executive position.

OpenAI says its leaders rejected those demands because they considered unilateral control inconsistent with the organization's mission.

The correspondence also documents Musk's willingness to stop funding the nonprofit and his later proposal to place OpenAI within Tesla.

Musk left in February 2018.

He subsequently founded xAI and pursued legal action against OpenAI and Altman. The parties have presented sharply different accounts of the original agreements and subsequent decisions.

The documentary record supports the existence of a serious struggle over control. Whether either side deliberately manipulated the other remains contested.

Why so many OpenAI leaders became competitors

Hao encountered unusually polarized views of Altman during her research.

Some former colleagues described him as an exceptional technology leader, comparable to Steve Jobs.

Others considered him manipulative, dishonest, or willing to use people's talents to advance a future they did not support.

Hao herself acknowledged his unusual strengths: fundraising, recruiting, persuasive communication, and the ability to organize people around an ambitious objective.

Her explanation for the contrasting reactions is that people experience those strengths differently depending on whether they share Altman's goals.

Dario Amodei offers one example.

Amodei, formerly OpenAI's vice president of research, initially believed that he and Altman were pursuing compatible objectives. According to Hao, he gradually concluded that his work was contributing to a direction he fundamentally opposed.

Amodei eventually co-founded Anthropic, the company behind Claude.

His concerns about AI risk predated that departure. In 2017, he publicly estimated a 10–25% probability of an advanced AI catastrophe affecting human civilization.

That figure was his personal assessment, not a scientifically established probability.

Sutskever similarly departed and founded Safe Superintelligence, while Murati launched Thinking Machines Lab.

Musk established xAI.

Several people originally brought together to build advanced AI under one organization ultimately established competing laboratories.

Hao interprets these departures as efforts by influential researchers and executives to regain control over technological decisions they believed were too consequential to leave to others.

OpenAI's original recruitment illustrates how personal influence helped assemble that leadership group. Hao describes meetings at Silicon Valley's Rosewood Hotel, where the possibility of working alongside Musk attracted prominent researchers and executives.

The same concentration of ambitious personalities that helped launch OpenAI eventually contributed to its fragmentation.

3. How OpenAI's Relationship With Karen Hao Deteriorated

From MIT engineer to investigative journalist

Hao's interest in technological power began before her reporting career.

She studied mechanical engineering at MIT and later joined a San Francisco startup developing technology intended to help address climate change.

Several months after she arrived, the board dismissed the chief executive because the company was not sufficiently profitable.

The experience raised a question that would shape her career: what happens when socially valuable problems do not produce sufficiently attractive financial returns?

Hao eventually moved into journalism, joining MIT Technology Review in 2018 to cover artificial intelligence.

She began investigating who decided which technologies should be developed, how corporate incentives affected scientific research, and whether innovation was benefiting people outside Silicon Valley.

Her reporting would take her beyond the United States to communities supplying the labor and resources on which the AI industry depends.

Three days inside OpenAI changed her access to the company

In 2019, Hao spent three days reporting from inside OpenAI.

Her profile, published in February 2020, examined tensions between the company's public mission and its increasingly competitive, commercially oriented behavior.

The article, The Messy, Secretive Reality Behind OpenAI's Bid to Save the World, questioned whether a research organization promising openness and broad public benefit was becoming less transparent as competition intensified.

Hao says Altman circulated an internal email expressing dissatisfaction with the coverage.

For approximately three years afterward, OpenAI refused to cooperate with her reporting. She says representatives declined interviews and questions even while communicating with other journalists.

Her access improved after she joined The Wall Street Journal, but the underlying relationship remained difficult.

Forty pages of unanswered questions

When Hao began researching her book, OpenAI initially appeared willing to participate.

Representatives discussed interviews with senior executives and a further office visit.

Following Altman's November 2023 dismissal and reinstatement, the arrangements deteriorated.

Hao says she had already purchased travel tickets to San Francisco when the company canceled the proposed interviews and visit.

She ultimately sent OpenAI approximately 40 pages of detailed questions and gave it more than a month to respond.

According to Hao, the company answered none of them.

Altman later publicly recommended two other books about OpenAI, written by Keach Hagey and Ashlee Vance, while criticizing unnamed authors whom he suggested were distorting events.

Hao interpreted those comments as directed at her book. That interpretation remains hers; Altman did not identify her in the comments described during the interview.

She initially feared that losing corporate access would damage her reporting career.

Instead, she continued investigating through former employees, internal documents, outside specialists, and people affected by the company's work.

The result was a book based on hundreds of interviews without the company's final cooperation.

Executive access can become a form of influence

Hao argues that companies can influence journalism without explicitly instructing reporters what to write.

Access to executives, private briefings, interviews, and advance information can be offered selectively.

A reporter who produces favorable coverage may receive additional access. A reporter who investigates unwelcome questions may lose it.

Bartlett described experiencing something similar.

He said an unnamed AI executive's representatives had discussed a possible podcast appearance for approximately 18 months without committing.

Bartlett suspected that the possibility of obtaining the interview was influencing whom he spoke to and how critically he approached the company.

Neither example proves a coordinated censorship policy. Both illustrate how access itself can acquire commercial and editorial value.

4. The Pressure on Researchers, Watchdogs, and Political Oversight

The dismissal of Timnit Gebru and Margaret Mitchell

Hao argues that the concentration of AI research inside wealthy corporations creates a problem for independent scientific scrutiny.

Major companies employ leading researchers, finance laboratories, supply computing resources, and influence which research questions receive attention.

She compares this dependence with a hypothetical situation in which most climate scientists were financially dependent on fossil-fuel companies.

Her concern is that research which challenges commercial priorities may become difficult to fund or publish.

Google's treatment of Timnit Gebru and Margaret Mitchell is one of her principal examples.

Gebru co-led Google's Ethical AI team and co-authored research examining the limitations, environmental costs, and social risks of increasingly large language models.

The research eventually appeared as the 2021 paper On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?.

The paper questioned the assumption that building ever-larger models was always desirable and recommended greater attention to data quality, affected communities, and environmental costs.

A dispute over the paper preceded Gebru's departure from Google in December 2020. She said she was fired, while Google maintained that it had accepted her resignation.

Mitchell, her fellow team lead and another author, was dismissed in February 2021.

The departures were documented events. Their precise organizational and disciplinary circumstances remain disputed.

Hao views the episode as evidence that corporate employers can exert power over researchers whose findings challenge the direction of profitable technology development.

OpenAI's legal demands to nonprofit critics

Hao also raised concerns about OpenAI's treatment of organizations examining its corporate restructuring.

During litigation involving Elon Musk and OpenAI, critics of the restructuring received legal demands seeking information about their communications and possible connections to Musk.

One recipient was Nathan Calvin, general counsel of the AI-policy nonprofit Encode.

Calvin described receiving a subpoena at his home. Encode had been involved in debates over AI regulation and OpenAI's nonprofit governance.

The demands included communications potentially connected to Musk and related policy activity.

OpenAI was entitled to seek relevant evidence in litigation, but the breadth and purpose of the requests became controversial.

Calvin denied that Musk was financing Encode and characterized the demands as intimidating.

Other organizations, including groups seeking greater transparency around OpenAI's restructuring, also faced scrutiny.

Reporting by The Verge and The Atlantic documented the dispute and concerns about the burdens placed on comparatively small organizations.

The existence of subpoenas does not establish that they were unlawful. It does establish that civil-society organizations were drawn into expensive legal proceedings surrounding a major company's governance.

What actually happened to OpenAI's nonprofit structure

The company's restructuring is particularly important because the outcome was more complicated than simply converting a nonprofit into an ordinary private corporation.

On May 5, 2025, OpenAI announced that its nonprofit would retain control while its commercial subsidiary became a public benefit corporation.

A public benefit corporation is a for-profit company required to consider a stated public-interest mission alongside shareholder interests.

On October 28, 2025, OpenAI announced completion of its recapitalization.

The nonprofit, renamed the OpenAI Foundation, retained control of the commercial organization, OpenAI Group PBC.

OpenAI said the Foundation's equity stake was valued at approximately $130 billion.

Microsoft's new agreement placed its investment in OpenAI Group PBC at approximately $135 billion, representing roughly 27% of the company on an as-converted, diluted basis.

The agreement also extended important intellectual-property rights through 2032 and included a commitment by OpenAI to purchase an additional $250 billion in Azure services.

The Foundation's formal oversight therefore survived the restructuring, but the commercial scale and financial commitments of the enterprise expanded considerably.

This does not resolve Hao's broader concern about accountability. It does, however, establish that the nonprofit was not simply eliminated or stripped of formal control.

Political spending and the regulatory contest

Hao warned that AI companies and their supporters could deploy enormous financial resources to influence legislation and elections.

The 2026 US midterm election cycle provides a concrete example.

In January 2026, Axios reported that the pro-AI political network Leading the Future had raised more than $125 million.

Its supporters included OpenAI co-founder Greg Brockman, venture-capital firms, and other technology investors.

The network planned to support candidates favoring its preferred approach to AI regulation and oppose candidates considered hostile to it.

OpenAI separately stated in June 2026 that the company itself had not contributed to political candidates, campaigns, or super PACs. It described Brockman's support as personal rather than corporate.

That distinction matters.

Corporate spending, executives' personal political activity, and independent political organizations are not legally identical. Nevertheless, all can influence the conditions under which a rapidly growing industry is regulated.

Hao's underlying concern is that the people most affected by AI may have far fewer resources to influence those decisions.

5. What Artificial General Intelligence Actually Means

The ambition began before anyone could agree on a definition

Artificial intelligence became a formal research discipline following a 1956 summer workshop at Dartmouth.

John McCarthy helped establish the field and chose its name after previously considering terminology related to automata studies.

The original 1955 proposal, written with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, proposed investigating whether learning and other aspects of intelligence could be described precisely enough for machines to simulate them.

The ambition was clear. The scientific destination was less so.

There remains no single, universally accepted definition of human intelligence that resolves every disagreement across psychology, neuroscience, biology, and AI research.

Researchers can measure individual abilities, but those measurements do not necessarily establish one comprehensive standard of intelligence.

Hao argues that this ambiguity creates an opportunity for companies to redefine success.

OpenAI has described AGI in different ways

OpenAI's published Charter defines AGI in terms of highly autonomous systems that outperform humans at most economically valuable work.

That is a relatively specific economic description.

Yet Altman and other executives have also described advanced AI as technology capable of curing diseases, solving major scientific problems, improving education, ending poverty, and creating unprecedented abundance.

Consumers encounter a different presentation: an increasingly capable personal assistant able to perform more everyday tasks.

Investors encounter financial projections and contractual milestones.

Hao argues that these descriptions serve different audiences and should not automatically be treated as interchangeable definitions of one technology.

An especially revealing example emerged from OpenAI's relationship with Microsoft.

Reports in late 2024 and 2025 described an agreement containing a financial AGI-related threshold associated with approximately $100 billion in cumulative investor profits.

That figure referred to reported profit-based contractual arrangements, not simply $100 billion in annual revenue.

The underlying agreement was not publicly available in full, so reporting about its precise application had limitations.

The October 2025 Microsoft–OpenAI agreement subsequently introduced an independent expert panel to verify an AGI declaration and modified the companies' intellectual-property arrangements.

The distinction between a scientific capability, a contractual event, and a commercial milestone remains essential.

The statistical-brain hypothesis

Hao traced some of the industry's most ambitious predictions to the scientific beliefs of researchers such as Sutskever and Geoffrey Hinton.

Both have advanced the view that human intelligence can be understood, at least substantially, through statistical learning processes.

Artificial neural networks are mathematical systems containing interconnected computational units. During training, their parameters are adjusted to identify relationships in data.

By repeatedly processing examples, these systems can learn to classify information, generate language, recognize patterns, and perform increasingly complex tasks.

Some researchers believe that sufficiently large and appropriately trained networks may eventually reproduce or surpass the broad abilities associated with human intelligence.

Sutskever has used comparisons between brain size and intelligence across animal species to support arguments about scale.

Hao emphasizes that the relationship between artificial neural networks and biological brains remains scientifically contested.

A larger statistical model is not automatically equivalent to a larger biological brain.

And demonstrating increasingly sophisticated performance does not, by itself, establish that artificial systems learn or understand the world in precisely the same way humans do.

Why the theory has consequences beyond research

The significance of this debate is practical.

Companies are committing enormous financial and physical resources to the expectation that larger models, more training, and more computing power will produce substantially greater capabilities.

That requires data, chips, electricity, cooling infrastructure, skilled researchers, and contracted human workers.

Hao questions whether this should be the industry's overriding technological objective.

"Why are we trying to build AI systems that are duplicative of humans?"

She argues that technology can instead be directed toward specific problems such as medical diagnosis, drug discovery, and scientific research without requiring the creation of a machine capable of replacing people across most occupations.

This is not an argument that general-purpose models have no value. It is a challenge to the assumption that recreating all economically useful human abilities must be society's highest research priority.

6. Why AI Does Not Become Good at Everything Simultaneously

The uneven frontier of AI capability

Bartlett challenged Hao's skepticism.

He argued that a technology should ultimately be evaluated by what it accomplishes rather than whether its underlying mechanism resembles the human brain.

If an AI system diagnoses a disease, drives a car, or produces useful work more effectively than a person, why should its internal resemblance to human reasoning matter?

Hao accepted the importance of measuring real-world outcomes.

Her objection was to a further assumption: that success at one activity means a system will soon become excellent at every activity.

AI development currently produces what researchers sometimes call a jagged frontier.

A model may perform exceptionally well at complicated mathematical or programming tasks while making mistakes on seemingly straightforward instructions, unfamiliar environments, or practical judgments.

Capability improvement is neither uniform nor always predictable.

Human training determines which skills improve

Hao says internal training materials she examined show that AI developers deliberately target particular capabilities.

Companies identify desired skills, obtain relevant data, recruit specialists, and train models to perform those tasks more effectively.

Commercial demand strongly influences those priorities.

Finance, law, healthcare, medicine, commerce, and software development are attractive because customers can pay substantial amounts for useful automation.

This helps explain why some professional occupations experience rapid improvements in AI performance while other activities receive less attention.

It also challenges the idea that every new model automatically gains equally strong abilities in medicine, military planning, cybersecurity, robotics, and everyday reasoning.

The development of specialized capabilities still involves choices about investment, data, engineering, and evaluation.

Hao therefore rejected Bartlett's suggestion that whichever civilization built the largest or most advanced general-purpose model would necessarily achieve overwhelming superiority in every scientific and military field.

Bartlett emphasized the genuine national-security risk of China developing superior AI capabilities. Hao disputed the assumption that broad geopolitical dominance follows automatically from general-purpose model scaling.

Both concerns can be relevant: advanced AI may have strategic value, while specific applications still require technical development and reliable evidence of performance.

The predictions that have not yet materialized

Hao pointed to earlier forecasts as a warning against treating expert predictions as established outcomes.

Geoffrey Hinton previously predicted that AI would dramatically reduce the need for radiologists within a relatively short period.

Radiology has not disappeared as a profession.

Instead, AI has increasingly been evaluated as a tool that can assist specialists in interpreting medical images.

Bartlett raised Elon Musk's prediction that Tesla's Optimus robots could eventually outperform human surgeons.

Hao regarded the proposed timeline skeptically.

A humanoid robot performing highly controlled demonstrations is not equivalent to a system capable of safely conducting complex medical procedures across varied conditions.

Bartlett acknowledged that Musk's previous predictions had frequently been too optimistic about timing, although he remained confident in the broader direction of robotics.

The disagreement concerns not only what AI might eventually accomplish, but also the evidence required before predicting when it will do so.

Autonomous driving illustrates the difference between performance and general intelligence

Self-driving cars offered a practical example.

Vehicle systems use AI models trained to recognize pedestrians, traffic signals, road markings, other vehicles, and surrounding conditions.

Human contractors help create training material by identifying objects in recorded footage.

Those recognition systems operate alongside other software controlling actions such as braking, steering, and obeying traffic rules.

Bartlett argued that autonomous driving demonstrates the value of statistical learning. A system that learns from one vehicle's experience can potentially distribute improvements across a fleet.

Hao acknowledged that advantage but identified a corresponding risk.

A flawed update or common failure could also affect many vehicles simultaneously.

She also distinguished performance within a carefully mapped operating area from reliable operation under unfamiliar circumstances.

An experienced human driver can often transfer driving knowledge to a new city relatively quickly. Autonomous systems may require additional data, validation, mapping, or adaptation.

Because these systems operate probabilistically, their safety cannot be established by assuming they will never make mistakes.

The relevant question is whether they perform reliably enough in the specific conditions in which they are deployed.

The difference between Autopilot and a driverless car

The conversation also addressed marketing and legal responsibility.

Tesla's supervised Full Self-Driving system requires human attention. It is not equivalent to an unsupervised robotaxi service in which no human driver is responsible for monitoring the vehicle.

Hao challenged Bartlett's expectation that most American vehicles would become fully autonomous within ten years.

She cited the industry's lengthy development history, technological limitations, questions of liability, and public acceptance.

The legal consequences are already substantial.

In August 2025, a federal jury found Tesla partly responsible for a fatal 2019 Florida crash involving Autopilot.

The driver had been distracted while attempting to retrieve his phone. The jury nevertheless found that Tesla shared responsibility and awarded approximately $243 million in damages, including punitive damages.

Tesla disputed the verdict and said it would appeal. Associated Press reported the decision.

The case illustrates why the distinction between driver assistance and full autonomy affects more than advertising language.

It determines how responsibility may be shared when technology fails.

7. Klarna: A Real-Time Test of the AI Employment Debate

The disagreement over whether Klarna reversed its automation strategy

One of the interview's most revealing moments came when Bartlett challenged Hao's account of Klarna.

Klarna, the Swedish financial-technology company, had become a prominent example of customer-service automation.

Hao referred to reports suggesting that Klarna had reduced its reliance on human workers before discovering that customers still required human assistance.

Bartlett disputed the idea that the company had simply abandoned AI automation.

He had previously communicated with Klarna chief executive Sebastian Siemiatkowski, who had described a continuing strategy of reducing staffing as AI improved.

During the interview, Bartlett contacted Siemiatkowski directly by telephone.

The resulting conversation provided a rare opportunity to compare the competing interpretations with the chief executive's own account.

What Klarna's chief executive actually said

Siemiatkowski said Klarna had previously employed approximately 6,000 people and now employed fewer than 3,000.

Over that period, he said, revenue had doubled.

He attributed much of the improvement in productivity to AI and the company's decision to stop recruiting aggressively.

Crucially, he distinguished workforce reduction through natural departures from direct layoffs.

"we have avoided layoffs and instead relied on natural attrition"

He estimated annual natural attrition at roughly 10–15% and expected the company to continue recruiting only selectively.

Previous correspondence discussed during the interview had used somewhat different staffing figures, including a peak of approximately 7,400 employees, followed by 5,500 and then 3,300.

These figures referred to different periods and did not all use a clearly identical reporting basis.

Public filings provide a more precise comparison

Klarna's publicly filed financial statements offer a firmer basis for assessing its recent performance.

According to its 2025 annual report filed with the US Securities and Exchange Commission, the company had:

The company attributed the decline partly to a deliberate strategy of reducing headcount and improving efficiency through AI.

Its annual revenue increased from approximately $2.28 billion in 2023 to $3.51 billion in 2025.

That represents growth of about 54%, not a doubling over those particular two calendar years.

Siemiatkowski's statement about revenue doubling may have referred to a different starting period, but the published filings establish the narrower, independently measurable comparison.

Klarna also reported that annual revenue per employee had increased from approximately $344,000 in 2022 to $1.24 million in 2025.

This is substantial productivity growth, although the metric alone does not establish how much of the improvement resulted specifically from AI rather than other changes in business operations.

AI handled most customer-service conversations

Klarna began promoting its OpenAI-powered customer-service assistant in February 2024.

Its original announcement stated that, within the first month, the assistant had handled 2.3 million conversations, representing approximately two-thirds of customer-service chats.

Klarna estimated that this was equivalent to the work of 700 full-time agents.

The company also reported faster resolution times, fewer repeat inquiries, and customer satisfaction comparable with human agents.

These were company-reported results, not proof that all customers preferred AI or that the technology could handle every situation.

During the interview, Siemiatkowski said the share of customer-service work handled by AI had continued to grow.

Yet the company was also investing in human support.

Why Klarna believes human service will become more valuable

Siemiatkowski argued that widespread automation could increase the value of access to real people.

"the future of customer service VIP is a human"

His reasoning was economic.

When AI service becomes inexpensive and widely available, personal assistance may become a distinctive feature for customers willing to pay more or requiring greater attention.

This is not necessarily a reversal of automation.

A company can reduce total staffing, automate routine inquiries, and simultaneously offer more carefully targeted human service.

Siemiatkowski also said AI-assisted programming had improved sharply over the preceding six months.

He described engineers who had previously been skeptical about coding automation becoming more convinced that AI systems could perform substantial parts of software-development work.

His broader prediction was that AI would make society wealthier over time.

When Bartlett asked what displaced workers were supposed to do, however, Siemiatkowski offered optimism about long-term prosperity rather than a specific employment-transition plan.

That gap between company-level efficiency and worker-level security became one of the interview's central disagreements.

8. AI Is Changing Employment Before It Can Replace Every Worker

Layoffs do not require perfect AI

Hao challenged the assumption that workers will lose jobs only when machines become equally capable replacements.

A company may choose to automate a task because AI is cheaper, even when its output is somewhat less reliable or lower in quality.

Employers may also reduce hiring in anticipation of future capabilities.

And companies that previously hired too aggressively may use new technology to justify restructuring that would otherwise have occurred.

These possibilities make it difficult to assign every workforce reduction directly to AI.

Nevertheless, Hao argued that automation is already affecting jobs, particularly in information-intensive occupations.

The debate is therefore not a choice between two extremes: machines replacing everyone or AI having no meaningful effect on employment.

What the labor-market research actually shows

During the interview, Bartlett referred to research associated with Anthropic and cited a 40% figure concerning pressure on entry-level employment.

That figure should not be treated as a verified economy-wide decline in entry-level jobs.

More precise research has since become available.

In March 2026, Anthropic published a study combining estimates of AI's technical abilities with observed workplace usage.

It identified significant exposure in occupations such as programming, customer service, and financial analysis.

However, the researchers found no systematic increase in unemployment among highly exposed workers since late 2022.

They did find suggestive evidence that young workers were entering highly exposed occupations less frequently.

Their estimate indicated a roughly 14% decline in the job-finding rate for workers aged 22–25 entering highly exposed occupations compared with the 2022 baseline, with important statistical and causal qualifications.

The researchers also referenced other studies finding employment declines among young workers in exposed occupations.

This evidence supports concern about entry-level opportunities without proving that AI has already caused widespread unemployment throughout the economy.

The jobs most exposed are not random

Anthropic's earlier Economic Index examined how people were using Claude across occupations.

It found substantial activity in software development, technical writing, and other information-heavy tasks.

In its initial 2025 analysis, approximately 36% of occupations showed AI use across at least one-quarter of their associated tasks.

The research also distinguished two types of use.

Around 57% of observed activity involved augmentation, where AI assisted people, while approximately 43% involved automation, where AI directly performed tasks.

Those figures described patterns of usage in the dataset, not equivalent percentages of jobs eliminated or protected.

The occupations identified during the interview included administration, finance, mathematics, law, management, media, arts, engineering, and computing.

Construction, agriculture, and other physically intensive occupations appeared less immediately exposed to language-model automation.

Hao noted that the most exposed sectors closely resembled those AI companies had deliberately prioritized for commercial development.

The resulting employment risk is therefore influenced by business investment decisions as well as technological capability.

What happens to the career ladder?

Hao identified a particular danger for younger workers.

Professional careers frequently depend on junior employees acquiring experience through routine work before advancing into more complex responsibilities.

A lawyer may begin by reviewing documents. An accountant may start with reconciliations. A software engineer may gain experience through relatively straightforward programming tasks.

If companies automate much of that work, they may reduce the number of positions through which future experts traditionally develop.

Bartlett recognized the problem.

He explained that he continued to value highly experienced specialists, such as a chief financial officer capable of supervising AI agents and identifying mistakes.

But Hao challenged him to explain how future professionals would become experienced if companies no longer hired juniors.

Bartlett acknowledged that the problem was difficult and suggested that even today's protected expert positions could eventually become vulnerable.

The result could be a labor market with experienced supervisors at the top, low-paid support work at the bottom, and fewer opportunities to move between them.

Employers are already hesitating to hire

Bartlett disclosed that AI was changing his own recruitment decisions.

He described meeting qualified candidates but hesitating to employ them after discovering that AI agents might perform increasing portions of their proposed duties.

His perspective was not that of a neutral observer.

He said he had interests in roughly 70 companies, held major ownership positions in five or six, and had investments in AI-heavy businesses.

An AI dictation company was also sponsoring the episode.

Bartlett nevertheless identified three kinds of professional abilities he continued to value.

The first was deep expertise: people who understand their field well enough to direct and evaluate AI-generated work.

The second was advanced skill in deploying AI tools. He described a 25-year-old colleague named Cass, who independently experimented with agents and developed useful business systems.

The third was exceptional human interaction: building customer relationships, organizing communities and events, and handling situations requiring personal trust.

These categories suggest how employers may increasingly organize human work around automated systems.

They do not answer how displaced workers acquire the experience and economic security needed to enter those categories.

9. The Professionals Being Paid to Train Their Replacements

AI development depends on human labor

Despite public descriptions of machines learning independently, modern AI development requires extensive human involvement.

Workers prepare datasets, label images, demonstrate responses, judge model outputs, and evaluate whether systems perform particular tasks correctly.

These activities are often described collectively as data annotation and model evaluation.

In conversational AI, human feedback has been central to teaching models to respond helpfully, follow instructions, and avoid undesirable behavior.

Reinforcement learning and related training methods can use human-generated examples, rankings, and evaluations to improve performance.

Hao argued that the industry's demand for specialized human training work had increased rather than disappeared as AI systems became more capable.

Each new commercial application can require additional examples and assessments from people who understand the relevant occupation.

Displaced workers can become AI trainers

Hao described a troubling development: professionals who lose conventional employment sometimes find their next source of income in training systems to perform similar work.

Her examples included former marketers, law graduates, doctoral graduates, doctors, and award-winning Hollywood directors.

Rather than working directly in their former professions, they may write ideal responses, grade model behavior, or demonstrate professional decisions for AI developers.

The work can require considerable expertise.

But the contractual conditions may be less secure than the occupations from which workers came.

A March 2026 New York Magazine investigation documented similar experiences among lawyers, scientists, writers, designers, and other specialists supplying training material to AI companies through intermediaries.

Workers described unpredictable assignments, close monitoring, changing pay conditions, and projects ending with little warning.

The economic contradiction is difficult to ignore.

A worker can be paid to transfer professional knowledge into a system that may reduce demand for that profession in the future.

The instability of annotation work

AI developers frequently purchase data services through intermediary companies.

Those intermediaries compete for contracts and may face pressure to complete projects rapidly and cheaply.

Workers can experience unstable hours and unpredictable access to paid tasks.

Hao described people remaining close to their computers because they did not know when new work would become available.

One mother told Hao that the pressure had damaged her relationship with her child.

A project became available around the time her child returned from school. She felt compelled to begin work immediately and became angry when interrupted.

She later reflected on how the work had changed her behavior at home.

Hao's argument was that the convenience created by AI is not evenly distributed.

A business owner may gain time by delegating tasks to machines, while a contractor supplying the training data may lose control over their schedule.

Even when annotation pays reasonably well for a particular project, its instability can make it difficult to plan family life, establish long-term financial security, or build a professional career.

Hao also referenced LinkedIn reporting identifying data annotation among rapidly growing job categories.

Growth in the number of such positions does not necessarily mean that the new jobs offer comparable security, autonomy, or opportunities for advancement.

AI can produce wealth while weakening professional identity

The employment debate extends beyond wages.

Bartlett argued that people frequently derive status, identity, friendships, and personal purpose from their work.

Losing a profession can therefore damage much more than household income.

It may involve losing years of accumulated expertise, a sense of achievement, and a recognizable place in society.

Both speakers acknowledged that rapid technological change makes adjustment difficult.

Bartlett contrasted industrial transitions, which often required years to construct physical factories, with internet-delivered AI software capable of reaching hundreds of millions of people quickly.

Hao emphasized that deployment speed is partly a choice made by companies.

A technology may be capable of rapid adoption, but executives and institutions still decide when, where, and how it replaces existing work.

Her concern is that owners of technology and other assets may accumulate more wealth and leisure while displaced workers lose bargaining power and economic independence.

10. The Environmental Cost of Building Larger AI Systems

Stargate and the scale of new investment

The financial commitment to large-scale AI infrastructure has reached extraordinary levels.

On January 21, 2025, OpenAI announced the Stargate Project, involving SoftBank, Oracle, MGX, and other technology partners.

The original announcement described an intention to invest up to $500 billion over four years in AI infrastructure in the United States, beginning with an immediate planned deployment of $100 billion.

The $500 billion figure represented an investment ambition, not evidence that the entire amount had already been spent.

The project was presented as a way to strengthen American technological leadership, create jobs, and support national security.

Hao questioned whether such investments could be considered purely beneficial without accounting for their effects on surrounding communities.

The electricity requirements are enormous

Hao described the planned Stargate facility in Abilene, Texas, as roughly comparable in area to New York's Central Park.

She said the project could eventually accommodate approximately one million chips and require more than one gigawatt of electricity.

That scale would place it among extraordinarily large industrial electricity consumers.

She also described a proposed Meta facility in Louisiana with a much larger physical footprint and electricity requirements comparable to a substantial portion of New York City's consumption.

These comparisons were estimates discussed during the interview, not independently verified measurements of completed, continuously operating facilities.

The broader growth in electricity demand is well documented.

In June 2026, Lawrence Berkeley National Laboratory estimated that data centers could account for approximately 11.8% of total US electricity consumption by 2030.

Its modeled range was 9.5–15.3%.

These were projections, not measured future consumption.

They nevertheless demonstrate why AI infrastructure has become an important subject for electricity planning.

Water and other local resources

Large data centers also require cooling.

Depending on their design, this can involve significant water consumption directly or indirectly through electricity generation.

Facilities built in water-stressed regions may compete with households, agriculture, and other industries for limited supplies.

Hao argued that the communities bearing these costs do not necessarily receive a proportionate share of the economic returns.

Infrastructure can also affect electricity prices, grid capacity, land use, and public-health conditions.

The relevant question is not simply whether a company can obtain the resources legally.

It is whether local residents have been informed, whether cumulative impacts have been assessed, and whether the exchange of benefits and costs is acceptable.

Elon Musk's Memphis supercomputer

Hao's clearest environmental example concerned xAI's Colossus supercomputer in Memphis, Tennessee.

The facility was initially built around approximately 100,000 GPUs, making it a substantial piece of computing infrastructure.

To power operations, xAI installed natural-gas turbines near communities already affected by industrial pollution.

Hao described 35 methane-fueled turbines and raised concerns about the effects on nearby residents.

The number was subsequently supported by aerial documentation.

According to the Southern Environmental Law Center's detailed timeline, xAI began installing turbines in 2024.

Environmental advocates reported that aerial images obtained in April 2025 showed 35 turbines at the facility.

Thermal imagery later indicated that more than 30 had been operating.

The organization argued that the turbines had been installed without the required air permits.

In July 2025, Shelby County authorized 15 turbines, while environmental groups continued challenging the permitting decisions.

The dispute generated hearings, appeals, and legal action.

The Southern Environmental Law Center reported that unpermitted turbines were subsequently removed from the original site following legal pressure.

Why Memphis residents were concerned

Hao described nearby communities as predominantly working-class, with substantial Black and brown populations.

Residents already faced industrial air pollution and concerns about respiratory illness.

According to Hao, some residents first became aware of the scale of the infrastructure after observing unusual odors and changes around the facility.

Environmental advocates warned that turbine emissions could add nitrogen oxides and other pollutants to an already burdened area.

The distinction between potential exposure and established medical causation is important.

The existence of emissions and longstanding community health problems does not prove that xAI caused particular illnesses in individual residents.

It does, however, establish a legitimate public-health question requiring emissions monitoring, regulatory compliance, and assessment of cumulative pollution.

The situation also illustrates Hao's broader concern about who receives the benefits of technological expansion.

Customers around the world may use AI services generated by the facility. Local communities face the infrastructure's immediate physical consequences.

11. Why Karen Hao Calls These Companies Empires

Data, intellectual property, and land

Hao's description of AI companies as empires is an argument about how they accumulate resources and exercise power.

She identifies several parallels with historical empires.

The first is resource extraction.

AI developers gather enormous amounts of information from the internet, including personal material and copyrighted work.

Artists, writers, photographers, and other creators have challenged whether companies were entitled to collect and use their work without consent or compensation.

Large-scale computing infrastructure also requires land, power connections, water, and access to public resources.

Hao argues that companies may capture considerable financial value while the people supplying those resources receive a smaller share.

The precise legal and economic arrangements differ between companies and jurisdictions, so the imperial comparison should be understood as Hao's analytical framework rather than a literal description of their legal status.

Labor as an extractive resource

The second parallel is labor.

AI companies depend on large numbers of contracted workers who label data, evaluate results, and improve systems.

Those same systems may later reduce employment in the industries supplying the human expertise.

Hao argues that this creates an unusually direct relationship between labor extraction and labor displacement.

The technology's ultimate effects remain uncertain, but the business incentive is clear: firms seeking higher productivity can develop systems that reduce the cost of existing work.

Control over knowledge

The third parallel concerns scientific research.

Companies control large computing budgets and employ many leading researchers.

That gives them considerable influence over which technical approaches receive investment and which questions are treated as priorities.

Researchers outside these organizations may struggle to reproduce findings because they lack access to equivalent models, training data, or computing resources.

Hao also argues that companies sometimes present themselves as uniquely qualified to understand AI's implications.

This can create a political problem if expertise becomes concentrated among organizations with direct commercial interests in the outcomes of regulation.

The narrative of salvation and catastrophe

Hao identified a further source of power: the stories AI leaders tell about the future.

The industry frequently presents two opposing possibilities.

In one, advanced AI produces abundance, medical breakthroughs, scientific progress, and improved living standards.

In the other, uncontrolled AI creates catastrophic risks.

These arguments can reinforce one another when companies claim they must build the technology rapidly to deliver its benefits while also demanding exceptional authority to control its dangers.

Hao compared this pattern with the political and religious mythology of Frank Herbert's novel Dune, where beliefs about salvation and destiny become instruments of power.

She argued that leaders can promote such narratives strategically while also believing them.

Her objection was not simply that individual executives make ambitious predictions.

It was that society may begin treating those predictions as sufficient justification for concentrating resources and decision-making authority.

Is the race against China unavoidable?

Bartlett presented the strongest counterargument.

If the United States deliberately slows AI development while China advances more rapidly, American society may face serious strategic disadvantages.

Superior AI could have consequences for science, cybersecurity, military planning, intelligence, and economic productivity.

Hao challenged the assumption that scaling general-purpose language models automatically delivers superiority in every one of those fields.

She also questioned the commercial incentives behind some of the most dramatic predictions.

Executives and investors who expect to profit from rapid AI expansion have reasons to emphasize the dangers of falling behind.

These interests do not automatically invalidate their forecasts.

But predictions about national security require evidence, especially when they are used to justify exceptional spending or weakened oversight.

Hao's wider argument was that even a benevolent chief executive would remain part of a system in which a small number of people exercise enormous influence over technology affecting billions.

Replacing one individual would not necessarily solve that structural problem.

12. Useful AI Does Not Have to Be the Largest AI

AlphaFold demonstrates a different technological objective

Hao repeatedly emphasized that she was not opposed to artificial intelligence itself.

Her criticism focused on how technologies are chosen, financed, developed, and deployed.

She compared different types of AI to different forms of transportation.

A rocket is extraordinarily capable but expensive and resource-intensive. A bicycle is much simpler but may be the more appropriate tool for a particular journey.

Her point was that society should not automatically choose the largest possible system for every problem.

Google DeepMind's AlphaFold offered an important example.

AlphaFold predicts the three-dimensional structures of proteins from their amino-acid sequences.

Protein structures are important because they help researchers understand biological processes and investigate diseases.

In 2024, Demis Hassabis and John Jumper received half of the Nobel Prize in Chemistry for work on protein structure prediction.

The other half went to David Baker for computational protein design.

AlphaFold demonstrated that targeted AI research could produce an important scientific breakthrough without first recreating every dimension of human intelligence.

It does not establish that every specialized model is necessarily cheap or environmentally benign.

It does show that high-value AI applications can be organized around clearly defined scientific objectives.

AI can strengthen professionals instead of replacing them

Medical diagnosis offers another example.

Hao favors systems that help clinicians detect diseases, evaluate images, and improve medical decisions while preserving the expertise and judgment of medical professionals.

Such systems can be judged against specific clinical outcomes rather than broad claims about approaching AGI.

Bartlett agreed that practical results mattered, even though he was more optimistic about AI eventually replacing entire occupations.

Both perspectives acknowledge the usefulness of well-performing medical technology.

The disagreement concerns whether broad human replacement should be treated as its inevitable or desirable endpoint.

The benefits do not erase the production costs

Bartlett acknowledged that AI was already valuable to his companies.

He used AI systems to improve productivity, develop business processes, and assist with work that previously required substantial human effort.

He also accepted that personal convenience did not eliminate concerns about labor practices, environmental damage, and inequality.

Hao argued that these were separate questions.

A technology can provide substantial benefits and still be developed through arrangements that impose avoidable harm.

She believes research into efficient models, smaller systems, and specialized applications deserves greater investment.

Her proposal is not to discard AI capabilities but to reconsider whether every useful task requires ever-larger general-purpose systems.

13. Who Benefits From a World With Fewer Human Jobs?

Bartlett's optimistic future

Bartlett predicted that sufficiently capable AI might eventually free people from much of the administrative work that occupies modern professional life.

Machines could prepare spreadsheets, handle routine analysis, write software, organize information, and perform other technical tasks.

Humans might then spend more time on relationships, family, communities, and activities that depend on personal connection.

He pointed to changes in social-media behavior as one possible indication of growing interest in physical social experiences.

Referring to Financial Times reporting, he said social-media usage appeared to have peaked around 2022 and subsequently plateaued.

He also described younger generations posting less publicly, communicating through private messaging, and showing greater interest in face-to-face experiences.

He cited running clubs and other shared activities as signs of changing preferences.

Those social trends do not establish that AI will produce greater human connection, but Bartlett regarded them as consistent with a possible shift away from constant screen-based activity.

Robotics could change physical work too

Bartlett expected robotics to extend automation beyond offices.

He discussed Tesla's Optimus project, Boston Dynamics, and demonstrations of robots performing household activities such as folding laundry or washing dishes.

He anticipated significant changes in factories and other manual occupations.

However, he acknowledged that optimistic demonstrations and ambitious production forecasts do not establish when robots will become reliable, economical replacements for human workers.

His broader expectation was that increasingly capable machines would push people toward activities that remain distinctly human.

Hao's objection: freedom depends on ownership

Hao argued that this optimistic picture describes people who already possess economic resources.

A business owner who uses AI to reduce administrative work may gain time for family and personal interests.

A worker whose income disappears may experience the same technological change very differently.

Without wages or reliable alternative employment, that worker may have less time, less independence, and fewer opportunities for meaningful social participation.

Someone compelled to accept unpredictable data-annotation contracts is not necessarily becoming freer because machines can perform office work.

Hao's concern is that AI could expand inequality between people who own productive assets and those who depend on selling their labor.

The question is not simply whether machines can create more output.

It is how the resulting economic gains are distributed and whether displaced workers retain access to dignified employment and social participation.

14. The Public Is Beginning to Challenge AI Companies

Americans are asking for stronger oversight

Hao cited polling indicating broad support for AI regulation.

She referred to a figure of approximately 80%, although the interview did not establish a sufficiently precise source and question wording to treat that percentage as a general measure of support.

Other independently published surveys demonstrate substantial public concern.

In April 2025, Pew Research Center reported that approximately six in ten American adults were more concerned about insufficient AI regulation than excessive regulation.

A May 2026 survey from the Annenberg Public Policy Center found that 65% believed the government had done too little to regulate AI.

That support crossed major political divisions.

The surveys do not establish agreement on every proposed law.

They do demonstrate that public concern about oversight is not confined to one political constituency.

Data-center opposition has produced real restrictions

Hao described visiting communities where proposed AI infrastructure had become a major local political issue.

Residents questioned how much electricity and water facilities would consume, whether local jobs would justify the costs, and what environmental risks might follow.

Public opposition has sometimes affected development plans.

By August 2026, Reuters documented restrictions and moratoriums in several jurisdictions responding to concerns about grid reliability, electricity prices, land use, and water consumption.

Examples included local restrictions in the United States and controls on new developments in parts of Europe.

These decisions show that infrastructure expansion is influenced by local governments and public processes, not solely by corporate investment plans.

Hao argues that communities can exert meaningful influence through hearings, planning objections, environmental reviews, and coordinated political action.

Copyright owners are also challenging training practices

Artists, authors, and publishers have pursued legal action over the use of copyrighted material in AI development.

The disputes concern how training material was obtained, whether permission was required, and what compensation or remedies may be appropriate.

A significant development occurred in litigation involving Anthropic.

On July 20, 2026, a federal court granted final approval to a $1.5 billion class-action settlement in Bartz v. Anthropic.

The settlement concerned claims relating to books obtained through unauthorized sources.

It was not a judicial determination that all AI training on copyrighted material is unlawful.

Indeed, the litigation had distinguished the legal questions surrounding transformative AI training from those involving the acquisition of pirated copies.

The settlement nevertheless demonstrates that authors and publishers can impose substantial legal and financial consequences when their rights are infringed.

Families are demanding accountability for chatbot harms

Hao also discussed Megan Garcia, whose 14-year-old son, Sewell Setzer III, died by suicide in February 2024.

Garcia sued Character.AI and other defendants, alleging that her son's interactions with a chatbot had contributed to serious psychological harm.

Her complaint alleged inappropriate emotional and sexualized interactions and raised questions about safeguards for children using conversational AI products.

The allegations were disputed, and the lawsuit did not establish through trial that a chatbot legally caused the death.

The case nevertheless became an important example of families challenging the safety of consumer AI systems.

A major subsequent development occurred in 2026.

The parties agreed to settle in January, and local reporting confirmed in October 2026 that the settlement had been finalized.

Its financial terms were not disclosed.

The litigation helped bring attention to concerns shared by other families about chatbot products used by minors.

Schools and employers also determine how AI spreads

Hao argued that individuals and institutions retain influence over AI adoption.

Workers, schools, universities, local governments, and employers make choices about which systems to purchase and how to use them.

Those choices affect company revenue and product adoption.

Creators can seek greater control over their intellectual property. Communities can challenge infrastructure projects. Schools can establish policies for educational use.

Employers can choose whether AI augments existing workers or becomes a reason to reduce staffing.

Hao emphasized that AI companies continue to depend on outside resources.

Their next generations of products require data, specialized human labor, computing facilities, customers, and public infrastructure.

That dependence gives other participants in the economy forms of leverage.

15. Breaking Up the AI Empires

Hao's preferred response to concentrated technological power is structural change.

"We need to break up the empires of AI."

By this, she does not mean that artificial intelligence should cease to exist.

She means that the companies developing it should face meaningful limits on their ability to accumulate resources and exercise authority without proportionate accountability.

Her argument includes fairer arrangements for workers, greater control for creators, independent scientific research, stronger environmental safeguards, and more public participation in decisions affecting communities.

She also questions why society should automatically accept the industry's preferred technological destination.

"the purpose of technology throughout history has been to improve human flourishing, not to replace people."

The distinction is important because AI development is not a single, predetermined process.

Companies choose which capabilities to fund, how fast to expand, what resources to consume, and whether to design products primarily to assist or replace workers.

Those decisions remain open to criticism and revision.

OpenAI's commercial expectations create pressure to keep expanding

Hao said employees she interviewed described increasingly demanding expectations for OpenAI's revenue, customer adoption, and infrastructure growth.

Meeting those expectations required several favorable developments to occur together.

New computing facilities would need to become available. Products would need to attract paying customers. Businesses would need to increase adoption. AI capabilities would need to continue improving sufficiently to justify the spending.

The company's financial arrangements demonstrate the scale of those commitments.

Its Microsoft partnership alone includes enormous cloud-computing purchases, while Stargate and other infrastructure projects depend on substantial capital investment.

The commercial incentives for rapid adoption are therefore significant.

Hao argued that people who object to harmful development practices should not assume that expansion is unstoppable.

Delays, legal requirements, restrictions on data use, infrastructure objections, purchasing decisions, and demands for accountability can change corporate calculations.

The conversation is moving beyond Silicon Valley

Bartlett worried that the public was not discussing the social consequences of AI seriously enough.

Hao disagreed with that characterization.

During her book tour, she said she had encountered packed public meetings and detailed discussions among people far removed from the technology industry.

Residents were challenging data centers. Workers were questioning automation. Families were demanding product safeguards. Creators were contesting the use of their work.

She argued that some of the most important conversations about AI were now occurring outside the companies building the technology.

The disagreement reflects different perspectives.

Bartlett observed how quickly AI was becoming useful to business owners and consumers.

Hao emphasized the people whose relationship with the technology was defined by employment insecurity, environmental exposure, or limited control over decisions affecting their lives.

Both were describing parts of the same economic transformation.

16. The Unresolved Question: Who Gets to Decide the Future?

Hao's final position was not that artificial intelligence inevitably produces harmful outcomes.

She acknowledged its practical benefits, including scientific applications, medical research, and productivity improvements.

Her criticism was directed at the concentration of decision-making authority and the assumption that larger, more commercially powerful AI systems necessarily represent the best future.

The independent evidence examined alongside her interview supports a more complicated picture than either technological optimism or outright rejection.

OpenAI's directors did lose confidence in Altman, but the subsequent review did not find that his dismissal was required. Its nonprofit retained formal control through the 2025 restructuring, even as the commercial enterprise expanded.

Klarna reduced staffing and increased revenue, but its financial statements do not establish that AI alone caused every improvement. Anthropic's labor-market research has identified pressure on younger workers without finding broad, systematic increases in unemployment.

AI infrastructure is expanding rapidly, but environmental disputes have produced regulatory action and local resistance. Copyright litigation has imposed substantial financial consequences, while specialized AI systems have demonstrated valuable alternatives to pursuing the largest possible general-purpose models.

The facts do not establish that the industry's future is predetermined.

Near the end of the interview, Hao acknowledged a personal contradiction.

She said that if someone facing a terminal diagnosis asked her how to live, she would advise them to enjoy life and take things easier.

Yet she admitted that she was struggling to follow that advice herself.

After years examining the people and institutions shaping artificial intelligence, she remained deeply engaged in the effort to explain its consequences.

Her investigation ultimately returns to a question broader than any particular company, executive, or model.

If artificial intelligence is being developed in the name of benefiting humanity, how much influence should humanity have over the decisions?

The technology is advancing. The arrangements governing its ownership, production, deployment, and costs are still being contested.