AI Whistleblower: We Are Being Gaslit By AI Companies, They’re Hiding The Truth! - Karen Hao

Karen Hao spent more than eight years investigating the AI industry for her book Empire of AI: Dreams and Nightmares in Sam Altman’s OpenAI. She conducted more than 300 interviews with over 250 people, including more than 90 current or former OpenAI employees and executives, to reconstruct the company's first decade and examine its worldwide consequences.

Hao entered journalism after becoming disillusioned with Silicon Valley's priorities. She studied mechanical engineering at MIT and joined a San Francisco startup developing climate-change technology. Months later, its board fired the CEO over profitability, convincing her that commercially successful innovation and public benefit were not necessarily aligned.

Her AI reporting began at MIT Technology Review in 2018. She wanted to investigate who decides which technologies get built, how money and ideology influence those decisions, and whether technological progress actually benefits people outside Silicon Valley.

Artificial intelligence became a formal research discipline in 1956 at Dartmouth. John McCarthy chose the name after previously considering “Automata Studies.” Hao argues that the eventual name tied the field to recreating human intelligence before scientists had established what human intelligence actually means.

There is still no universally accepted scientific definition of human intelligence. Hao argues that this makes the industry's ambition to achieve artificial general intelligence (AGI) difficult to measure and allows companies to change what success means.

OpenAI has presented different definitions of AGI to different audiences. Hao says Sam Altman describes it to lawmakers as technology capable of curing cancer, solving climate change and eliminating poverty; to consumers as an exceptional digital assistant; and in Microsoft's investment arrangements as a system associated with $100 billion in revenue.

OpenAI's public definition of AGI focuses on economically valuable work. Its website has described AGI as highly autonomous systems outperforming humans at most economically valuable tasks. Hao argues that this differs materially from the broader humanitarian promises used to promote the technology.

Hao believes AGI's changing definition serves commercial and political purposes. Different descriptions can attract investment, persuade consumers, reassure lawmakers or discourage regulation without requiring the company to commit to one measurable technological destination.

Sam Altman's 2015 warnings about existential AI risk closely resembled Elon Musk's public statements. Before OpenAI's announcement, Altman wrote that superhuman machine intelligence was potentially humanity's greatest existential threat. Hao says he was simultaneously trying to recruit Musk as a co-founder.

Altman had previously emphasized engineered viruses as a major threat. Hao interprets his 2015 shift toward catastrophic AI risk as an effort to align with Musk, who had publicly compared developing advanced AI to summoning a demon.

Musk ultimately co-founded and financially supported OpenAI but later came to believe Altman had manipulated him. Hao says documents disclosed in their legal dispute provide additional context for Musk's resentment, although the interpretation of Altman's intentions remains contested.

Musk was initially favored to lead OpenAI's proposed for-profit operation. According to emails discussed by Hao, Ilya Sutskever and Greg Brockman initially supported Musk becoming CEO when the organization considered creating a commercial entity alongside its nonprofit.

Altman privately persuaded Brockman to reconsider Musk's leadership. Hao reports that Altman appealed to their longstanding friendship and raised concerns that Musk's celebrity, outside pressures and unpredictability could make concentrating control of powerful AI in his hands dangerous.

Brockman subsequently convinced Sutskever to support Altman instead. When Musk discovered he would not become CEO, he withdrew. Hao presents this reversal as an important part of the breakdown between OpenAI's original leaders.

Opinions about Altman are unusually polarized among people who have worked with him. Hao encountered admirers who compared him with Steve Jobs and critics who described him as manipulative or dishonest. She says the divide largely reflected whether individuals supported his preferred technological future.

Altman's greatest organizational strength is his ability to persuade and mobilize resources. Hao credits him with exceptional storytelling, fundraising and recruitment abilities. Those aligned with his goals see these qualities as assets, while opponents may feel those same abilities are being used against their interests.

Dario Amodei's departure from OpenAI illustrates that conflict. Amodei, formerly OpenAI's vice president of research and now CEO of Anthropic, initially believed he and Altman wanted the same future. Hao says he gradually concluded that his work was advancing an agenda he fundamentally opposed.

Amodei was publicly warning about catastrophic AI outcomes as early as 2017. While at OpenAI, he estimated a 10–25% probability that advanced AI could go catastrophically wrong on the scale of human civilization. The figure was his personal assessment, not an established scientific probability.

Ilya Sutskever believed OpenAI had two inseparable responsibilities: achieve AGI and develop it safely. Hao says he eventually concluded that Altman's leadership was undermining both objectives, particularly through organizational disorder and competing internal teams.

Sutskever warned in 2019 that powerful AI might treat humans as humans treat animals. His analogy was that people do not necessarily hate animals but still build highways through their habitats without seeking permission. He feared autonomous AI could similarly disregard humanity's interests.

Sutskever and his mentor Geoffrey Hinton shared a hypothesis that human brains operate like statistical models. This belief supported their conviction that sufficiently large artificial neural networks might reproduce or surpass human intelligence. Hao stresses that the hypothesis is scientifically disputed.

Sutskever used comparisons between brain size and animal intelligence to support his scaling argument. In a research-conference presentation, he suggested that larger statistical models might eventually acquire greater intelligence than humans. Hao cautions that this depends on assumptions about the brain that have not been established.

The brain hypothesis has enormous practical consequences because companies organize their investments around it. Pursuing ever-larger models requires more training data, computational power, electricity, data centers and human workers, even though the underlying theory remains contested.

Hao questions why replicating or replacing humans became the industry's central ambition. She argues that developing systems for drug discovery, medical research and other specific benefits could be more useful than pursuing a machine capable of doing everything humans do.

Hao calls leading AI companies “empires” because of how they accumulate and exercise power. Her argument concerns their control over data, intellectual property, infrastructure, labor, scientific research and political decisions, rather than simply their size or profitability.

The first characteristic of these AI empires is resource extraction. Hao says companies claim or use individuals' data and creators' intellectual property to train models while securing enormous amounts of land and infrastructure for computing facilities.

The second characteristic is labor exploitation. Hao says AI development depends on hundreds of thousands of contracted workers worldwide, including Americans, whose labor teaches systems to perform tasks that may later be automated.

Labor displacement is partly a deliberate design choice. Hao argues that companies prioritize systems capable of replacing paid work, potentially weakening workers' bargaining power, rather than exclusively building technology that improves existing jobs.

The third characteristic is control over scientific knowledge. Hao argues that major AI companies fund or employ much of the research community, influencing which questions receive resources and how the technology's benefits and limitations are presented.

Industry dependence can discourage research that challenges commercial priorities. Hao compares it with a hypothetical world where most climate scientists depended financially on fossil-fuel companies, arguing that such concentrated sponsorship threatens independent evaluation.

Google's treatment of Timnit Gebru and Margaret Mitchell is Hao's example of research suppression. Gebru co-led Google's Ethical AI team and co-authored research examining harms associated with large language models. Following a dispute over that work, Gebru and fellow co-lead Mitchell were dismissed.

Hao says OpenAI also used legal demands against organizations criticizing its corporate restructuring. During the company's effort to convert its structure, a representative of a small watchdog nonprofit was served with a demand for communications potentially involving Elon Musk.

The watchdog groups were seeking transparency around OpenAI's conversion. Hao says organizations including the Midas Project wanted public scrutiny of a consequential change from the original nonprofit arrangement rather than a process conducted without meaningful debate.

OpenAI appeared to suspect Musk was secretly financing opposition to its restructuring. Hao says the targeted organizations were not funded by Musk and the watchdog representative responded that he had no relevant communications. She interprets the demands as potentially intimidating and an attempt to map relationships among critics.

Hao says OpenAI ultimately succeeded in its restructuring despite the opposition. Her broader concern is that the company used substantial legal and organizational resources while smaller civil-society groups struggled to secure transparency and public participation.

AI companies use competition with a threatening rival to justify concentrating power. Hao says OpenAI originally depicted Google as the dangerous commercial competitor that a benevolent nonprofit needed to defeat. China later became the more common geopolitical adversary in industry messaging.

Promises of extraordinary benefits and warnings of catastrophe reinforce each other. According to Hao, companies describe futures ranging from abundance and cures for major diseases to humanity's destruction, then argue that they must control development to ensure the favorable outcome.

Hao believes this rhetoric discourages democratic participation in technological decisions. If the technology is presented as too dangerous for ordinary oversight yet too valuable to delay, the companies developing it can demand exceptional freedom, resources and authority.

OpenAI's relationship with Hao deteriorated after her first major company profile. She spent three days embedded at OpenAI in 2019 for MIT Technology Review. When her profile appeared in 2020, Altman circulated an internal email expressing dissatisfaction with the coverage.

OpenAI subsequently refused to cooperate with Hao for approximately three years. She says representatives explicitly declined interviews, questions and opportunities to engage, even when they continued offering stories to other journalists.

Her access improved after she joined The Wall Street Journal. Hao continued requesting comments and interviews, and the company became more willing to communicate. She later left the newspaper to research her OpenAI book full-time.

OpenAI initially agreed to participate in Hao's book. Representatives discussed arranging executive interviews and another office visit, but the relationship changed after the board dismissed Altman in November 2023.

The company ultimately withdrew its participation after delaying arrangements. Hao says she had already purchased travel tickets for the planned San Francisco interviews when OpenAI informed her that neither the meetings nor the visit would proceed.

Hao sent OpenAI 40 pages of detailed requests for comment. She gave the company more than a month to respond to her reporting but says it did not answer any of the questions.

Altman publicly endorsed two competing books while criticizing unnamed authors. He identified books by Keach Hagey and Ashlee Vance as works OpenAI had supported and suggested some other writers were twisting events. Hao interpreted the criticism as directed at her forthcoming book.

Hao describes access to senior executives as a mechanism for influencing journalism. Companies can reward favorable reporters with interviews and information while withholding access from those who interview critics or pursue unwelcome stories.

Steven Bartlett described experiencing a similar pressure tactic. He said an unnamed AI executive's team had discussed a possible podcast appearance for around 18 months without committing, which he interpreted as an attempt to influence whom he interviewed and how critically he covered the company.

Losing OpenAI's cooperation eventually became an advantage for Hao's reporting independence. She initially feared it would damage her career, but said she learned to investigate without relying on company-approved access and still completed more than 300 interviews for the book.

Hao reconstructed Altman's 2023 dismissal using approximately six or seven sources close to the decision. Her account begins with Sutskever becoming increasingly concerned that Altman's behavior was damaging research, trust and decision-making inside OpenAI.

Sutskever cautiously approached independent board member Helen Toner. Hao says he initially revealed little because he feared retaliation if Altman discovered he was criticizing his leadership.

OpenAI's independent directors were intended to protect its public-interest mission. The governance structure distinguished independent members from people with financial stakes in the company, creating a mechanism to challenge decisions favoring commercial interests over the nonprofit's mission.

Sutskever and chief technology officer Mira Murati supplied board members with concerns and supporting documentation. Their discussions included emails, Slack messages and accounts of Altman allegedly giving conflicting messages to different teams and creating internal rivalries.

The executives argued that Altman's leadership had made collaboration increasingly difficult. Their concern was not merely that OpenAI was growing quickly, but that he was intensifying distrust and disorder at a company attempting to develop technology they considered extraordinarily consequential.

ChatGPT's unexpectedly successful launch produced substantial operational chaos. OpenAI had treated the GPT-3.5-based release as a research preview to collect data ahead of a more ambitious GPT-4 chatbot, rather than anticipating an immediate mass-market success.

OpenAI struggled with server outages and extremely rapid hiring following ChatGPT's launch. Hao says the company recruited aggressively, sometimes concluded that hires were mistakes, and dismissed employees so abruptly that colleagues discovered their departures through disappearing Slack accounts.

The independent directors judged OpenAI by a more demanding standard than an ordinary startup. Hao reports that they considered whether Altman's behavior would justify dismissal at a conventional technology company and concluded that the potential consequences of AGI required stronger safeguards.

An unusual arrangement involving the OpenAI Startup Fund deepened board concerns. Board member Adam D'Angelo reportedly heard questions about the fund's structure at a San Francisco gathering. After examining documents, directors discovered that what appeared to be OpenAI's fund was legally structured around Altman personally, according to Hao.

The Startup Fund became one of several examples of inconsistent information reaching directors. Hao says board members increasingly questioned differences between how Altman described organizational arrangements and what documentation showed.

The board decided to remove Altman quickly and secretly because members feared his persuasive influence. They believed advance notice would give him an opportunity to mobilize supporters and prevent the decision.

Microsoft received almost no advance warning of Altman's dismissal. Despite being one of OpenAI's most important investors, it was notified shortly before the action, leaving major stakeholders excluded from the decision.

The board's failure to prepare stakeholders helped bring Altman back. Investors, employees and other affected parties reacted strongly to being excluded, creating an effective campaign for his reinstatement within days.

Sutskever's concern that confronting Altman could endanger his position proved consequential. After Altman returned, Sutskever did not return to his former role, while Murati also subsequently left OpenAI.

Sutskever had privately questioned whether Altman should control AGI development. Bartlett cited Hao's account of Sutskever saying he did not believe Altman should have his finger on the button for AGI, illustrating how deeply the company's chief scientist had lost confidence in its leader.

OpenAI's original recruitment involved a gathering at the Rosewood Hotel in Silicon Valley. Altman used the possibility of meeting Elon Musk to attract researchers and executives, including Sutskever, Brockman and Amodei, to discussions about the new organization.

Many prominent figures recruited into OpenAI eventually departed after disagreements with Altman. Hao sees the repeated breakups as evidence that powerful researchers and entrepreneurs struggled to share control over their competing visions for AI.

Former OpenAI leaders established several major competitors. Musk founded xAI, Amodei established Anthropic, Sutskever launched Safe Superintelligence, and Murati founded Thinking Machines Lab. Hao interprets these companies as attempts by their founders to regain control over the direction of AI development.

Hao compares AI leadership culture with the mythology in Frank Herbert's Dune. She suggests executives consciously construct stories about technological salvation and danger, much as political figures in the novel use a messianic narrative to acquire power.

Executives can deliberately promote an AI narrative while also genuinely believing it. Hao argues that repeatedly communicating the same high-stakes predictions can blur the distinction between strategic persuasion and personal conviction.

Hao does not consider the morality of individual AI CEOs the central problem. Even replacing them with better-intentioned leaders would leave a structure in which a small group makes decisions affecting billions of people without those people having meaningful representation.

She argues that electoral politics alone cannot adequately restrain concentrated AI power. Hao says major companies have the financial resources to spend hundreds of millions of dollars influencing upcoming US midterm elections, opposing restrictive legislation and supporting rules advantageous to themselves.

Bartlett defended rapid American AI development on national-security grounds. He argued that China might otherwise develop superior scientific, military and cyber capabilities, leaving the United States dependent on or vulnerable to Chinese technology.

Hao rejected the premise that larger language models automatically produce greater general intelligence. She argued that several unproven assumptions must hold before superior AI research can be equated with military or geopolitical superiority.

Current AI systems have uneven rather than universally improving capabilities. Hao describes a “jagged frontier”: models can perform exceptionally well on selected tasks while struggling with others, even when those other tasks appear comparatively simple.

Companies choose which capabilities to improve instead of advancing every ability equally. Hao says internal training documents show developers selecting particular skills, gathering corresponding datasets and hiring contractors to teach models those tasks.

Commercial demand strongly influences the capabilities being developed. According to Hao, developers prioritize fields such as finance, law, medicine, healthcare and commerce because customers in those industries can pay for automation.

Military and cybersecurity improvements are not automatic consequences of model scaling. Hao acknowledges that AI has long been used in military systems but argues that stronger specialized capabilities require specific development choices rather than simply making general-purpose models larger.

Hao distinguished AI's statistical pattern recognition from human learning and adaptation. Humans can acquire knowledge independently and transfer experience between unfamiliar situations, whereas today's AI systems can require substantial additional training to operate in new environments.

Self-driving vehicles illustrate the limits of transferring learned capabilities. Hao says a system trained for one city may require extensive additional data and adaptation elsewhere, unlike an experienced human driver who can use general driving knowledge in a new location.

Shared AI learning creates both advantages and systemic risks. Bartlett emphasized that knowledge acquired by one vehicle or robot can be deployed across an entire fleet. Hao countered that the same process can spread an incorrect behavior or common failure across every system.

Bartlett argued that AI should ultimately be evaluated by its results, not whether it resembles a human brain. He cited his own experiences with autonomous driving and asked whether a machine that performs a task more safely or effectively really needs to satisfy a particular definition of intelligence.

Hao accepted the importance of measurable results but challenged the extrapolation to universal competence. She argued that demonstrating superior performance in selected tasks does not prove that models will progressively master everything humans can do.

Hao questioned the incentives behind leading executives' predictions of universal AI intelligence. She noted that figures including Musk, Altman, Amodei, Hinton and Zuckerberg promote versions of the scaling argument, while some have substantial commercial interests in its acceptance.

Musk's xAI built the Colossus supercomputer in Memphis around an initial 100,000-GPU installation. Hao presented the project as an example of the enormous capital and infrastructure being committed to developing larger AI models faster than competitors.

Musk predicted that AI-powered Optimus robots could outperform human surgeons within a few years. Bartlett raised the prediction as a test of whether Hao believed continued AI improvements would replace highly trained professionals.

Hao cited Geoffrey Hinton's earlier prediction about radiologists as a cautionary example. Hinton had predicted that AI would eliminate the need for radiologists within a period that has already passed, while radiology remains an active profession.

Hao favors medical AI that assists specialists rather than automatically replacing them. She cited research indicating that combining clinicians' expertise with AI can improve early detection and diagnosis of certain cancers.

Modern neural networks learn statistical relationships from training data. Hao explained that interconnected software nodes identify patterns that allow models to classify, predict and act without every behavior being programmed individually.

Autonomous vehicles require extensive human-labeled training material. Contractors identify pedestrians, traffic signals, lane markings and vehicles in recorded footage so AI models learn to recognize the surrounding environment.

Self-driving systems also use conventional programmed rules alongside AI. Recognition models identify objects, while other software governs responses such as stopping at red lights or avoiding pedestrians.

Probabilistic AI systems cannot be guaranteed to make zero errors. Hao argued that their statistical foundations create unavoidable uncertainty, making the safety question one of reliability and acceptable risk rather than perfect performance.

Hao acknowledged that autonomous driving can outperform humans in specific environments. She disputed extending those safety results to unfamiliar roads, countries and conditions where systems have not been adequately trained.

Hao rejected the prediction that most US vehicles would be autonomous within ten years. She pointed to the industry's long development history, technical limitations, public trust, legal liability and the comparatively small number of fully autonomous vehicles currently operating.

Liability remains a barrier to autonomous driving adoption. The conversation referenced a case involving a Tesla driver who looked down to retrieve a phone before a fatal collision, with both the driver and Tesla held responsible.

Bartlett distinguished Tesla's supervised Full Self-Driving from fully driverless vehicles. Hao emphasized that systems requiring a human to remain attentive are only partially autonomous, regardless of their marketing names.

Hao expects substantial employment disruption from AI but disputes the idea that layoffs simply reflect machines becoming more capable. Job losses also occur because executives choose cheaper automation, accept lower-quality outputs or reorganize their businesses in anticipation of future capabilities.

Some employers may replace workers before AI can fully perform their jobs. Hao argues that companies can accept a less capable system because its cost is sufficiently lower, or use AI as a justification for reducing previously excessive staffing.

Klarna became a disputed example of AI-driven workforce reduction. Hao initially referenced accounts that the company had reduced staffing and later needed human workers again. Bartlett challenged the simplified narrative and contacted CEO Sebastian Siemiatkowski directly during the interview.

Siemiatkowski had previously told Bartlett that Klarna was continuing to reduce staffing because of AI. In their earlier messages, he said the company had fallen from approximately 7,400 employees at its peak to 5,500 and then 3,300, with a target near 3,000.

Klarna reported that AI was handling roughly 70% of its customer-service conversations. Siemiatkowski attributed major efficiency improvements to automation and argued that AI was dramatically reducing the production cost of software.

During the live telephone call, Siemiatkowski said Klarna had roughly halved its workforce while doubling revenue. He described a reduction from approximately 6,000 employees to fewer than 3,000 over two to three years, presenting the results as evidence that AI enabled the company to accomplish more with fewer people.

Klarna's CEO explicitly said the reductions had occurred through natural attrition rather than layoffs. He said the company had largely stopped hiring, allowing annual employee departures of approximately 10–15% to reduce headcount without direct dismissals.

Klarna plans to continue hiring very selectively. Siemiatkowski expects the workforce to decline further through natural departures while AI handles increasing amounts of work.

Klarna is simultaneously investing in more human interaction for customers who value it. Siemiatkowski argued that when AI service becomes cheap and widely available, access to a real person can become a more valuable premium experience.

Siemiatkowski believes AI-assisted coding has recently undergone a major shift. He said improvements over the preceding six months had persuaded even formerly skeptical engineers that AI could perform much of the work traditionally done through manual programming.

Klarna's CEO remains optimistic about AI's long-term economic benefits despite acknowledging short-term employment disruption. He expects society to become wealthier but did not offer a detailed answer about how displaced workers would transition into suitable new roles.

Hao says the employment debate has become falsely divided between total automation and no meaningful disruption. Her position is that AI is already replacing some tasks and jobs, while business decisions, cost-cutting and user preferences determine the actual employment consequences.

The conversation cited weakening hiring in American white-collar occupations. Hao referred to employment reports indicating slower recruitment in professional fields, although she did not claim the entire slowdown could be attributed exclusively to AI.

Bartlett referenced an Anthropic analysis suggesting particularly severe pressure on entry-level employment. He cited a 40% reduction figure and a chart comparing current AI use with tasks models could potentially perform, emphasizing the gap between technical capability and actual adoption.

Anthropic's analysis concentrated potential disruption in information-intensive occupations. The categories discussed included office administration, finance, mathematics, law, management, media, arts, engineering and computer-related work, while physical occupations such as construction and agriculture appeared less immediately exposed.

Hao observed that these exposed occupations closely match the industries AI developers deliberately target. She considers the projected disruption partly a consequence of investment and product-development choices rather than an inevitable property of generalized intelligence.

People may continue paying for human expertise even when AI can perform the same technical task. Klarna's human-service strategy illustrated the importance of personal contact, trust and customer preferences alongside efficiency.

AI threatens the traditional career ladder by removing entry-level and intermediate work. Hao argues that automation may leave highly experienced professionals at the top and low-paid support work at the bottom, with fewer opportunities for people to acquire the experience needed to advance.

Displaced professionals are increasingly finding work training the systems that threaten their previous careers. Hao cited reporting on former marketers and other professionals becoming data annotators who teach AI to perform the jobs they once held.

The annotation workforce includes highly qualified people who cannot find suitable employment elsewhere. The conversation referenced law graduates, doctoral graduates, doctors and award-winning Hollywood directors performing such work to support themselves financially.

Data annotation is becoming a fast-growing occupational category. Hao cited LinkedIn reporting that included the work among its ten fastest-growing job categories, despite the uncertain conditions associated with many contracts.

Data annotation is essential to the conversational abilities of systems such as ChatGPT. Hao explained that large numbers of workers supplied examples of appropriate responses, helping transform language models from general text generators into systems capable of dialogue.

Reinforcement learning builds on human-provided examples and feedback. Annotators demonstrate or evaluate desirable behavior, while subsequent training processes use that information to improve particular model capabilities.

AI companies frequently purchase annotation services through intermediary contractors. Hao says firms compete for contracts from developers such as OpenAI and Google, creating strong incentives to complete labeling work quickly and cheaply.

The resulting work can be unstable, intensely monitored and poorly suited to normal family life. Workers may have to remain near their computers waiting for short-notice assignments because they do not know when another paying opportunity will appear.

Hao recounted a mother who felt unable to care properly for her child because of annotation work. When a project became available as her child returned from school, she began working frantically and shouted at the child for interrupting, later describing how the pressures had changed her behavior.

Hao argues that AI can restore time and freedom to business owners while removing both from workers. Owners can delegate tasks to machines, but the people supplying training data may experience less control over their schedules, earnings and personal lives.

Bartlett acknowledged that AI is already changing his own hiring decisions. He said he had encountered qualified candidates but hesitated to employ them after discovering that AI agents could increasingly perform the tasks for which he was recruiting.

Bartlett has commercial exposure to AI through his businesses and investments. He described involvement with roughly 70 companies, lead ownership positions in five or six, and investments in AI-heavy businesses, including an AI dictation company sponsoring the episode.

Bartlett identified deep professional expertise as one category of employees he still considers essential. His example was an experienced CFO who could supervise AI agents performing financial analysis because she understood the business and could judge their outputs.

He identified advanced AI-tool proficiency as a second valuable capability. He described a 25-year-old colleague, Cass, who experiments with AI agents independently and develops systems that improve business operations.

Bartlett's third valued category is people with exceptional real-world interpersonal abilities. He emphasized client relationships, organizing events, bringing communities together and maintaining workplace relationships as activities requiring human engagement.

Hao challenged Bartlett to explain how companies would develop future experts if they eliminated the junior positions that traditionally produce them. He acknowledged the difficulty and suggested that even today's relatively protected specialist roles could face further automation pressure.

Bartlett predicted that automation might eventually push people toward more meaningful human relationships. His optimistic scenario is that machines perform administrative and technical work while people devote more time to personal contact, community and activities machines cannot meaningfully replace.

Bartlett pointed to changing social-media behavior as evidence of a shift toward real-world interaction. He cited a Financial Times report suggesting social-media usage peaked around 2022 and said younger generations were posting less publicly, using private messaging more and showing greater interest in in-person experiences.

He also predicted major changes in physical labor through robotics. Referring to Optimus and companies such as Boston Dynamics, Bartlett anticipated increasingly capable robots performing domestic and factory work, although he acknowledged that Musk's previous timelines have often been inaccurate.

Hao argued that Bartlett's optimistic account mainly describes people who already possess power and resources. Business owners can choose to use automation for freedom and connection, while workers who lose employment may instead be pushed into insecure annotation jobs with less opportunity for human interaction.

The speed of AI adoption makes workforce adjustment especially difficult. Bartlett contrasted the years required to build factories during previous industrial transitions with AI software that can reach hundreds of millions of users rapidly through the internet.

Hao emphasized that the pace of disruption is not an uncontrollable natural force. Companies choose to compete aggressively, accelerate development and deploy automation, making them responsible for considering whether workers and institutions can adapt.

The conversation raised employment's importance beyond income. Bartlett described how professional displacement can destroy a person's sense of status, identity and purpose, potentially contributing to shame, depression and other serious personal consequences.

Hao's central economic concern is that AI could intensify existing inequality. People who already own businesses, assets and technology may gain money and leisure, while displaced workers face worse jobs, weakened bargaining power and reduced personal agency.

AI infrastructure creates additional burdens that are often concentrated in poorer communities. Hao says enormous computing facilities can increase electricity costs, strain power grids, consume water and introduce environmental hazards.

OpenAI's Stargate initiative was presented as a $500 billion AI infrastructure effort. Hao connected the initiative, announced near the beginning of Donald Trump's second administration, with the construction of major American AI computing facilities.

The proposed OpenAI facility in Abilene, Texas, was described as roughly the size of Central Park. Hao said it could contain approximately one million computer chips and require more than one gigawatt of electricity, equivalent to over 20% of New York City's power demand.

A separate Meta facility in Louisiana was described as even larger. Hao said its planned footprint was approximately four times that of the Abilene project, around one-fifth the size of Manhattan, with electricity demand equivalent to roughly half of New York City's average usage.

Large data centers can create competition for scarce freshwater. Water is needed for cooling and for some electricity-generation processes, and Hao cited communities already facing drought where residents must compete with industrial facilities for limited supplies.

Musk's Memphis Colossus facility became Hao's principal example of environmental injustice. She said the project used 35 methane gas turbines to power AI computing infrastructure near a predominantly working-class Black and brown community.

Hao said Memphis residents were not adequately informed about the project's environmental consequences. According to her account, some became aware of problems after noticing gas-like odors in their homes.

The Memphis community already faced longstanding air-quality and health problems. Hao said turbine emissions added to an existing burden of respiratory illness, childhood asthma and elevated lung-cancer risks, further illustrating the concentration of AI infrastructure's costs.

Hao rejected the idea that technological benefits excuse environmental and labor harms elsewhere. Her objection is that some communities may lose employment opportunities, pay higher utility bills and experience pollution while receiving relatively little of the wealth created.

Hao compares different AI technologies to different forms of transportation. Some systems resemble resource-intensive rockets, while others resemble efficient bicycles; she argues it is irrational to use the largest and most expensive technology for every problem.

Google DeepMind's AlphaFold is her example of a more targeted and efficient AI system. It predicts protein structures from amino-acid sequences, supporting disease research and drug discovery. The research associated with AlphaFold received the 2024 Nobel Prize in Chemistry.

Hao argues that specialized AI can deliver major benefits with much smaller datasets and computing requirements. Systems trained for well-defined scientific tasks can consume substantially fewer resources than enormous general-purpose models while producing valuable outcomes.

The industry's dependence on new training data has not disappeared as models become more capable. Hao challenged the suggestion that companies had already collected everything they needed, explaining that new generations require updated information and training for additional abilities.

Hao says demand for human data annotation has increased during her seven years investigating the sector. Rather than eliminating the need for human training work, expanding AI capabilities have created further requirements for specialized examples and evaluations.

The continuing dependence on outside data, labor and infrastructure gives the public leverage. Hao argues that companies cannot build and deploy their next generations of products entirely independently of people, communities and institutions that supply these resources.

Hao's preferred response is to break up concentrated AI power rather than abolish artificial intelligence. She wants companies to operate under fairer conditions and face meaningful constraints on extracting data, labor and public resources.

She defines an exploitative “empire” partly by its refusal to provide a fair exchange of value. Her concern is that companies can capture substantially greater benefits than the compensation or consideration they provide to workers, creators and affected communities.

Hao cited polling indicating that approximately 80% of Americans favor regulating AI. She presented this as evidence of unusually broad agreement that the industry needs stronger public accountability.

Opposition to data-center construction has become an important grassroots movement. Hao described protests in numerous communities, some of which have delayed projects or secured local restrictions and outright bans on new facilities.

Artists and writers are using intellectual-property litigation to challenge AI training practices. Hao sees these lawsuits as efforts to establish meaningful control over creative work and determine when companies must obtain permission or provide compensation.

Hao discussed Megan Garcia's legal campaign following the death of her son, Sewell Setzer III. The 14-year-old died by suicide after interactions with a Character.AI chatbot that his mother alleged involved sexual grooming. Garcia sued, helping draw attention to similar concerns raised by other families.

The legal actions by affected families have widened scrutiny of AI products used by children. Hao sees these lawsuits as examples of individuals challenging corporate practices that otherwise receive limited public oversight.

Hao urged people to examine where they personally supply resources or approval to AI companies. Relevant decisions include contributing personal or creative data, accepting a local data center, purchasing services, and determining AI adoption policies in workplaces and schools.

Withholding access to data is one form of public leverage. Hao pointed to creators attempting to establish restrictions through litigation, arguing that continued training depends on the industry's ability to obtain useful new material.

Communities can influence AI development through local infrastructure decisions. Public hearings, planning objections and coordinated opposition to data centers can affect the speed and cost at which computing facilities are built.

Schools and employers can influence adoption through their purchasing and usage policies. Hao suggested these institutions need not automatically accept AI products simply because the industry wants widespread deployment.

OpenAI faces unusually demanding revenue and adoption expectations, according to employees Hao recently interviewed. She said they described internal targets requiring many things to go exceptionally well, including rapid infrastructure expansion and extensive customer adoption.

Hao therefore urged people who oppose harmful practices not to make the industry's expansion effortless. Delaying or withholding consent, demanding accountability and choosing alternative technologies could influence business models that depend on continued rapid growth.

Hao believes research already points toward ways to preserve AI capabilities with fewer resources. She distinguished opposition to today's development incentives from opposition to the technology itself, arguing that better methods and smaller specialized systems deserve greater investment.

Bartlett acknowledged the tension between benefiting personally from AI and recognizing its social costs. He described the technology as extraordinarily useful in his companies and everyday life while accepting that those benefits do not eliminate concerns about environmental damage, inequality or worker displacement.

Hao argued that technological usefulness and harmful production practices are separate questions. Her position is that society can retain valuable AI capabilities while changing how systems are financed, trained, governed and deployed.

The speakers disagreed about how widely the public debate is already occurring. Bartlett worried that AI's social consequences were being discussed primarily within the industry, while Hao said she had encountered packed public meetings and serious discussions across dozens of cities during her book tour.

Hao said the most important conversations are increasingly happening outside Silicon Valley. She described local governments, affected communities, workers and ordinary citizens questioning who benefits from AI development and who bears its costs.

Hao's closing personal admission was that she struggles to follow the advice she would give someone facing a terminal diagnosis. She would encourage that person to enjoy life and take things easier, but acknowledged that she herself was not taking it easy while pursuing her reporting and advocacy.