Most Replayed Moment: How AI Companies Are Gaslighting You - Karen Hao

The host begins by examining how prominent AI researchers’ public statements evolved over time. He quotes Dario Amodei, formerly OpenAI’s vice president of research and later associated with Anthropic, who said in 2017 that he could not see any principled reason why an artificial general intelligence system could not destroy humanity. Amodei estimated the chance of something catastrophically going wrong on the scale of human civilization at between 10% and 25%. The conversation then turns to Ilya Sutskever, an OpenAI co-founder who later left the company. Karen Hao explains that Sutskever cared deeply about two goals: achieving AGI and achieving it safely. In her account, he came to believe that Sam Altman was undermining both by creating a chaotic internal environment, pitting teams against each other, and telling different people different things. Hao says this contributed to Sutskever’s effort to have Altman removed. Hao recalls interviewing Sutskever in 2019. At the time, he said that the future would be good for AIs regardless, and that it would be nice if it were also good for humans. He compared the likely relationship between autonomous AI and humans to the relationship between humans and animals: humans may love animals, but they do not ask animals for permission before building a highway. The implication was that a sufficiently powerful AI might treat humans as obstacles without actively hating them.

The host asks whether understanding the mechanism behind AI matters if the practical result is that systems can make videos, operate as agents, or perform people’s jobs. Hao answers that it matters because AI companies are making major decisions based on a particular hypothesis: that human brains are essentially statistical models, and that increasingly large statistical models will eventually produce systems with human-level or superhuman intelligence. Hao connects this belief to Sutskever’s research background and to his mentor, Geoffrey Hinton. At a major AI research conference, Sutskever presented a chart suggesting a roughly linear relationship between brain size and the intelligence of a species. If the brain is a statistical engine, Hao says, then building a larger digital statistical engine could logically produce a more intelligent entity. That framework also leads to the fear that humans could be treated as animals are treated by humans. Hao stresses that this is a scientific hypothesis, not an established fact. Critics argue that describing the brain as merely a statistical engine is reductive. She says the disagreement matters because companies are using the hypothesis to justify acquiring more data, constructing larger data centers, and employing more labor in pursuit of AGI. She also questions why the goal should be to build systems that duplicate and automate human capabilities at all, rather than focusing on applications such as accelerating drug discovery or improving healthcare.

When the host asks why these companies continue pursuing AGI, Hao says she believes they are driven by what she calls an “imperial agenda.” She uses the term “empires of AI” because, in her view, the empire metaphor captures both their scale and the methods they use. Hao identifies several parallels with historical empires. First, AI companies claim resources that do not belong to them, including individuals’ data and the intellectual property of artists, writers, and other creators. They also engage in a land grab for the sites needed to build massive computing facilities. Second, they rely on extensive labor exploitation, including hundreds of thousands of contracted workers around the world, while designing products intended to automate labor and potentially erode labor rights. Third, Hao argues that these companies monopolize knowledge production. They present themselves as the people who understand AI best and imply that public criticism comes from people who lack technical knowledge. At the same time, the industry employs or funds a large share of AI researchers, allowing it to influence which questions receive funding and attention. Hao points to Google’s treatment of Timnit Gebru and Margaret Mitchell: Gebru, Google’s ethical-AI team co-lead, co-authored research on harms from large language models, and Google fired her and later fired Mitchell after efforts to stop the paper’s publication. Hao says this shows how inconvenient research can be suppressed.

The host asks whether AI leaders genuinely believe the outcome will benefit everyone. Hao says the industry’s mythology requires two apparently opposing possibilities at once. Sam Altman has publicly described a worst-case scenario in which it is “lights out for everyone,” while presenting the best case as curing cancer, solving climate change, and creating abundance. Hao says Dario Amodei uses a similar structure: the worst case is catastrophic or existential harm, while the best case is mass human flourishing. According to Hao, the catastrophic scenario is not merely a warning; it helps companies justify keeping control of the technology. If AI could become an existential threat, they argue, then it must be developed by the companies and researchers who supposedly know how to manage it. Hao describes this as an anti-democratic approach because it excludes broad public participation while allowing the companies to control development at every stage. She says the identity of the supposed “bad empire” has shifted. In OpenAI’s early story, Google was portrayed as the dangerous, profit-driven corporation that might build the technology first. More recently, China is often presented as the rival that must be beaten. In both versions, the message is similar: the company’s own expansion is necessary because an opposing power might otherwise win. Hao argues that this narrative turns resource accumulation and control into a claimed public service.

The host raises an example involving a small watchdog nonprofit whose director had questioned OpenAI’s attempt to convert from a nonprofit into a for-profit company. During that period, the organization was trying to obtain more transparency and public debate around the conversion. Hao says OpenAI subpoenaed critics and, in one case, someone appeared at the director’s home to serve papers and seek every communication that might have involved Elon Musk. The apparent theory was that Musk was funding people who opposed the conversion, although Hao says the groups were not funded by him and the request produced no such documents. Hao also describes the use of access as leverage over journalists and other people with large communication platforms. She says technology companies can offer interviews, office visits, and executive access as a “carrot,” then withdraw it if a journalist speaks to critics or platforms views the company dislikes. The host confirms that an AI figure’s team had dangled a possible appearance on his show for roughly 18 months, apparently hoping that continued anticipation would shape how he covered the company. He says he would rather ask genuinely curious questions and let viewers judge the answers. Hao’s own relationship with OpenAI illustrates the cost of refusing that access game. After her 2020 MIT Technology Review profile made the leadership unhappy, OpenAI refused to participate in her work for about three years. When she later began writing her book, the company initially offered interviews, but withdrew after the board fired Altman. Hao sent more than 40 pages of requests for comment and gave the company over a month to respond; it did not answer any of them. She ultimately says the closed door strengthened her ability to report based on the evidence rather than protecting access.