Empire of AI

Through my reporting, I’ve come to understand two things: Artificial intelligence is a technology that takes many forms. It is in fact a multitude of technologies that shape-shift and evolve, not merely based on technical merit but with the ideological drives of the people who create them and the winds of hype and commercialization. While ChatGPT and other so-called large language models or generative AI applications have now taken the limelight, they are but one manifestation of AI, a manifestation that embodies a particular and remarkably narrow view about the way the world is and the way it should be. Nothing about this form of AI coming to the fore or even existing at all was inevitable; it was the culmination of thousands of subjective choices, made by the people who had the power to be in the decision-making room. In the same way, future generations of AI technologies are not predetermined. But the question of governance returns: Who will get to shape them?

But in the same way we once thought Facebook was merely a place for posting vacation pictures and connecting with long-lost elementary school friends, or for sparking positive and transformative social movements, there is more to the sleek, entrancing exterior than meets the eye. Under the hood, generative AI models are monstrosities, built from consuming previously unfathomable amounts of data, labor, computing power, and natural resources. GPT-4, the successor to the first ChatGPT, is, by one measure, reportedly over fifteen thousand times larger than its first generation, GPT-1, released five years earlier. The exploding human and material costs are settling onto wide swaths of society, especially the most vulnerable, people I met around the world, whether workers and rural residents in the Global North or impoverished communities in the Global South, all suffering new degrees of precarity. Rarely have they seen any “trickle-down” gains of this so-called technological revolution; the benefits of generative AI mostly accrue upward.

Over the years, I’ve found only one metaphor that encapsulates the nature of what these AI power players are: empires. During the long era of European colonialism, empires seized and extracted resources that were not their own and exploited the labor of the people they subjugated to mine, cultivate, and refine those resources for the empires’ enrichment. They projected racist, dehumanizing ideas of their own superiority and modernity to justify—and even entice the conquered into accepting—the invasion of sovereignty, the theft, and the subjugation. They justified their quest for power by the need to compete with other empires: In an arms race, all bets are off. All this ultimately served to entrench each empire’s power and to drive its expansion and progress. In the simplest terms, empires amassed extraordinary riches across space and time, through imposing a colonial world order, at great expense to everyone else.

As Baidu raced to develop its ChatGPT equivalent, employees working to advance AI technologies for drug discovery had to suspend their research and cede their computer chips to develop the chatbot instead. The current AI paradigm is also choking off alternative paths to AI development. The number of independent researchers not affiliated with or receiving funding from the tech industry has rapidly dwindled, diminishing the diversity of ideas in the field not tied to short-term commercial benefit. Companies themselves, which once invested in sprawling exploratory research, can no longer afford to do so under the weight of the generative AI development bill. Younger generations of scientists are falling in line with the new status quo to make themselves more employable. What was once unprecedented has become the norm.

There is a different way forward. Artificial intelligence doesn’t have to be what it is today. We don’t need to accept the logic of unprecedented scale and consumption to achieve advancement and progress. So much of what our society actually needs—better health care and education, clean air and clean water, a faster transition away from fossil fuels—can be assisted and advanced with, and sometimes even necessitates, significantly smaller AI models and a diversity of other approaches. AI alone won’t be enough, either: We’ll also need more social cohesion and global cooperation, some of the very things being challenged by the existing vision of AI development.

But the empires of AI won’t give up their power easily. The rest of us will need to wrest back control of this technology’s future. And we’re at a pivotal moment when that’s still possible. Just as empires of old eventually fell to more inclusive forms of governance, we, too, can shape the future of AI together. Policymakers can implement strong data privacy and transparency rules and update intellectual property protections to return people’s agency over their data and work. Human rights organizations can advance international labor norms and laws to give data labelers guaranteed wage minimums and humane working conditions as well as to shore up labor rights and guarantee access to dignified economic opportunities across all sectors and industries.

Both his media savvy and dealmaking, two pillars of his rise, rest on his remarkable ability to tell a good story. In this Altman is a natural. Even knowing as you watch him that his company would ultimately fail, you can’t help but be compelled by what he’s saying. He speaks with a casual ease about the singular positioning of his company. His startup is part of the grand, unstoppable trajectory of technology. Consumers and advertisers are clamoring for the service. Don’t bet against him—his success is inevitable.

Key to his formula, people say, is the combination of his remarkable listening skills, his willingness to help, and his ability to frame whatever he has to offer in terms of exactly what you want.

From both men, Altman also learned the importance of building relationships and creating “network effects” as an individual. “I’ve heard a lot of different theories about how things get done,” he wrote on his blog in 2013. “Here’s the best one: a combination of focus and personal connections. Charlie Rose said this to Paul Graham, who told it to me.” Altman would later add a third ingredient: self-belief. “For startups I think it’s really important to add this,” he said. “You actually have to believe you might do it.”

Over the following year, Donald Trump’s spectacular rise and win in the 2016 US presidential election would shock the left-leaning workforce of the tech industry into self-reflection. As upheaval ripped through companies like Meta and Google and techlash sentiment gripped the public, AI researchers, too, began to question whether the field had moved too quickly to yoke its technologies to corporate bottom lines. An accounting of the societal impacts of commercializing AI research returned an unsettling scorecard: Automated software being sold to the police, mortgage brokers, and credit lenders were entrenching racial, gender, and class discrimination. Algorithms running Facebook’s News Feed and YouTube’s recommendation systems had likely polarized the public, fueled misinformation and extremism, enabled election interference, and, most horrifying in the case of Facebook, precipitated ethnic cleansing in Myanmar.

Now here was a group of people—nine out of eleven of whom were white men—being showered in previously unheard-of amounts of money, speaking about the theoretical prospect of a bad superintelligence taking over the world, and proposing to counteract it by building a better superintelligence.

That night, Gebru drafted a scathing critique of what she’d observed in an anonymous open letter: the spectacle, the cultlike exaltation of AI celebrities, and, most of all, the overwhelming homogeneity of the people building and shaping such a consequential technology. This homogeneous culture was not only pushing away talented researchers but also leading to a dangerously narrow conception of AI and of who could benefit from the technology.

As Sutskever rallied researchers to give him their best ideas, Brockman threw himself into the work of developing the right organizational culture. Some years later, Brockman would recount to me his thinking. To prepare, he read every book he could find on ambitious science and technology undertakings in US history: the transcontinental railroad, Thomas Edison’s light bulb, the early network of computers that would lay the groundwork for the modern Web. He absorbed them like religious texts, searching for hints and guidance on how to design his own endeavor.

But this existential brand of AI safety, built on philosophical thought experiments, would soon come under fire as the AI research community awakened to the less apocalyptic and immediate real-world harms of AI. Around the same time Amodei published his paper, ProPublica published a groundbreaking investigation called “Machine Bias” that revealed algorithms were being used across the US criminal justice system in misguided attempts to predict future criminals, and those algorithms were classifying Black people as higher risk than white ones who had more extensive criminal records. The piece, and an overall souring on Big Tech post-2016 over the harms of social media, sparked a new wave of research reckoning with the harmful societal impacts of AI.

In early 2019, senior Microsoft leadership began coming through the OpenAI office. First came Kevin Scott, the tech giant’s excitable chief technology officer, who had followed OpenAI and grown particularly fond of the startup; then came Craig Mundie, a senior adviser to Nadella who had served on Microsoft leadership, including as its chief research and strategy officer, for over twenty years. Bill Gates also turned up, reserved and tight-lipped as usual, as he watched a series of demos. Most employees were left in the dark about Microsoft’s engagement. Altman told the small team working on the deal to keep knowledge of a possible investment limited.

My conversation with Brockman and Sutskever continued on in circles until we ran out the clock after forty-five minutes. I tried with little success to get more concrete details on what exactly they were trying to build—which by nature, they explained, they couldn’t know—and why then, if they couldn’t know, they were so confident it would be beneficial.

At one point, I tried a different approach, asking them instead to give examples of the downsides of the technology. This was a pillar of OpenAI’s founding mythology: The lab had to build good AGI before someone else built a bad one.

“Fundamentally, I want to work on AGI for the rest of my life,” Brockman told me.

He approached everything with this kind of intensity. He was hands on and detail oriented. He pulled long hours and all-nighters, consumed by the tasks in front of him. When he wasn’t working, he was still working to better himself. He read books about public speaking and negotiation, skills he believed would help him serve OpenAI’s mission. He was sheepish to admit that he also read books for fun—science fiction epics like The Three-Body Problem trilogy by Chinese writer Liu Cixin and Isaac Asimov’s Foundation series.

In their book Power and Progress, MIT economists and Nobel laureates Daron Acemoglu and Simon Johnson argue that every technology revolution must begin with a rallying ambition. It is the promise of a technology benefiting everyone that puts in motion the long journey of amassing enough talent and resources to turn it into a reality. After analyzing one thousand years of technology history, the authors conclude that technologies are not inevitable. The ability to advance them is driven by a collective belief that they are worth advancing. The irony is that for this very reason, new technologies rarely default to bringing widespread prosperity, the authors continue. Those who successfully rally for a technology’s creation are those who have the power and resources to do the rallying. As they turn their ideas into reality, the vision they impose—of what the technology is and whom it can benefit—is thus the vision of a narrow elite, imbued with all their blind spots and self-serving philosophies. Only through cataclysmic shifts in society or powerful organized resistance can a technology transform from enriching the few to lifting the many.

These two features of technology revolutions—their promise to deliver progress and their tendency instead to reverse it for people out of power, especially the most vulnerable—are perhaps truer than ever for the moment we now find ourselves in with artificial intelligence. Since its conception, the development and use of AI has been propelled by tantalizing dreams of modernity and shaped by a narrow elite with the money and influence to bring forth their conception of the technology. That conception is what has led to the exploding social, labor, and environmental costs that are playing out around the world today, particularly, as we’ll see, in many Global South countries, for which the consequences of their dispossession by historical empires still linger in delayed economic development and weaker political institutions. And yet, just like the South Carolina congressman, Silicon Valley has painted the experiences of those being exploited and harmed by the technology as happier because of it.

That tradition of anthropomorphizing continues to this day, aided by Hollywood tales combining the idea of “AI” with age-old depictions of human-made creations suddenly waking up. AI developers speak often about how their software “learns,” “reads,” or “creates” just like humans. Not only has this fed into a sense that current AI technologies are far more capable than they are, it has become a rhetorical tool for companies to avoid legal responsibility. Several artists and writers have sued AI developers for violating copyright laws by using their creative work—without their consent and without compensating them—to train AI systems. Developers have argued that doing so falls under fair use because it is no different from a human being “inspired” by others’ work. The omnipresent AI-to-human analogies have also fueled the sense that such software could become so capable that it surpasses us and comes to threaten our very existence. The fear of superintelligence is predicated on the idea that AI could somehow rise above us in the special quality that has made humans the planet’s superior species for tens of thousands of years.

Computer vision seeks to re-create our sight; natural language processing and generation, our ability to read and write; speech recognition and synthesis, our ability to hear and speak; and image and video generation, our creativity and imagination. As software for each of these capabilities has advanced, researchers have subsequently sought to combine them into so-called multimodal systems—systems that can “see” and “speak,” “hear” and “read.” That the technology is now threatening to replace large swaths of human workers is not by accident but by design.

Now, AI is the harbinger of the fourth industrial revolution. The keystone of the modern superpower. AGI, if ever reached, will solve climate change, enable affordable health care, provide equitable education. OpenAI is the poster child for this line of thought. It cannot say how the technology will deliver on these promises—only that the staggering price society needs to pay for what it is developing will someday be worth it.

Regardless of the answer, it has become clear that if they can, they do so inefficiently. Only with extraordinary amounts of data and computational power have neural networks even begun to have the kinds of behaviors that may suggest the emergence of either property. That said, one area where deep learning models really shine is how easy it is to commercialize them. You do not need perfectly accurate systems with reasoning capabilities to turn a handsome profit. Strong statistical pattern-matching and prediction go a long way in solving financially lucrative problems. The path to reaping a return, despite similarly expensive upfront investment, is also short and predictable, well suited to corporate planning cycles and the pace of quarterly earnings. Even better that such models can be spun up for a range of contexts without specialized domain knowledge, fitting for a tech giant’s expansive ambitions. Not to mention that deep learning affords the greatest competitive advantage to players with the most data.

That year, I stumbled across a Massachusetts-based, Harvard-incubated startup selling AI- powered headbands that said it could measure a student’s brain wave activity to tell a teacher whether or not the child was focused. The startup was piloting them in elementary schools in Colombia and China, in exchange for the rights to use their students’ data to advance the company’s technology.

“We have the first mover’s advantage,” a research scientist at the company had said at an education technology conference in 2017. “We’ll be able to build one of the largest brain wave databases in the world. All that data will help us improve our algorithms and therefore our products, creating a higher barrier to entry.”

A few months after I came across the startup, a data privacy outcry in China from parents horrified at their kids being turned into guinea pigs forced the company to pivot to a different application of its technology.

But the story left me with an uneasy feeling that the successful backlash was an anomaly, and the company’s original approach—to go to countries eager to embrace the promise of technology for finding data donors and product testers—was in fact a trend.

As I recounted this worry to a colleague, she introduced me to a phrase that had already been coined for the phenomenon: “data colonialism.”

This has led to dangerous outcomes. In March 2018, a self-driving Uber killed forty-nine-year-old Elaine Herzberg in Tempe, Arizona, in the first ever recorded incident of an autonomous vehicle causing a pedestrian fatality. Investigations found that the car’s deep learning model simply didn’t register Herzberg as a person. Experts concluded that it was because she was pushing a bicycle loaded with shopping bags across the road outside the designated crosswalk—the textbook definition of an edge-case scenario. Six years later, in April 2024, the National Highway Traffic Safety Administration found that Tesla’s Autopilot had been involved in more than two hundred crashes, including fourteen fatalities, in which the deep learning–based system failed to register and react to its surroundings and the driver failed to take over in time to override it. The fallible and inscrutable statistical patterns of neural networks can also turn into a security vulnerability. In 2019, white hat hackers tricked a Tesla in self-driving mode into veering into an incoming lane of traffic.

Their modern-day nemesis was Gary Marcus, a professor emeritus of psychology and neural science at New York University, who would testify in Congress next to Sam Altman in May 2023. Four years earlier, Marcus coauthored a book called Rebooting AI, asserting that these issues were inherent to deep learning. Forever stuck in the realm of correlations, neural networks would never, with any amount of data or compute, be able to understand causal relationships—why things are the way they are—and thus perform causal reasoning. This critical part of human cognition is why humans need only learn the rules of the road in one city to be able to drive proficiently in many others, Marcus argued.

In reality, the analogies to intelligence are once again anthropomorphizing and exaggerating the capabilities of the technology. While Hinton and other deep learning absolutists predicted that the shortfalls of neural networks compared with humans would go away at sufficient scale, the challenges have in fact persisted and, by many accounts, only gotten worse.

Microsoft’s Nadella has similarly called Bing’s AI chat “search, just better”—a tool “to be able to get to the right answers.” Even the term hallucinations is subtly misleading. It suggests that the bad behavior is an aberration, a bug, when it’s actually a feature of the probabilistic pattern-matching mechanics of neural networks.

As ChatGPT swept the world by storm in early 2023, a Chinese AI researcher would share with me a clear-eyed analysis that unraveled OpenAI’s inevitability argument. What OpenAI did never could have happened anywhere but Silicon Valley, he said. In China, which rivals the US in AI talent, no team of researchers and engineers, no matter how impressive, would get $1 billion, let alone ten times more, to develop a massively expensive technology without an articulated vision of exactly what it would look like and what it would be good for.

Through the course of my reporting, I would come to conclude something even more startling. Not even in Silicon Valley did other companies and investors move until after ChatGPT to funnel unqualified sums into scaling. That included Google and DeepMind, OpenAI’s original rival. It was specifically OpenAI, with its billionaire origins, unique ideological bent, and Altman’s singular drive, network, and fundraising talent, that created a ripe combination for its particular vision to emerge and take over. “I get the sense that Sam is the most ambitious person on the planet,” a former employee says. In other words, everything OpenAI did was the opposite of inevitable; the explosive global costs of its massive deep learning models, and the perilous race it sparked across the industry to scale such models to planetary limits, could only have ever arisen from the one place it actually did.

In the coming months, Amodei and Altman would clash over how and when to release GPT-3; Altman would win out, pushing the model into the world on an accelerated timeline. Years before ChatGPT, these two decisions—the one to explode GPT-3’s size and the one to quickly release it—would change the course of AI development. It would set off a rapid acceleration of AI advancement, sparking fierce competition between companies and countries. It would fuel an unprecedented expansion of surveillance capitalism and labor exploitation. It would, by virtue of the sheer resources required, consolidate the development of the technology to a degree never seen before, locking out the rest of the world from participating. It would accelerate the vicious cycle of universities, unable to compete, losing PhD students and professors to industry, atrophying independent academic research, and spelling the beginning of the end of accountability. It would amplify the environmental impacts of AI to an extent that, in the absence of transparency or regulation, neither external experts nor governments have been able to fully tabulate to this day.

In a 2023 paper, Abeba Birhane and her coauthors would introduce the concept of “hate scaling laws” to critique the premise of training deep learning models on unfiltered data, or what they called “data-swamps.” They analyzed two publicly available image-and-text datasets used to train open-source image generators, LAION-400M and LAION-2B-en, both pulled from Common Crawl, with four hundred million and two billion images, respectively. They showed that the amount of hateful and abusive content scaled with the size of the dataset and exacerbated the discriminatory behaviors of the models trained on them. Models trained on the two billion images, for example, were five times more likely than models trained on the four hundred million images to label Black male faces as criminals. Later that year, a Stanford study analyzing LAION-5B, a dataset with five billion images used to train Stable Diffusion, would discover it contained thousands of images of verified and suspected child sexual abuse.

Early in his career, Altman observed that new CEOs only succeeded if they “refounded” the company. He did this with conviction at YC when he inherited the presidency. He created new programs, including a new fund, to expand the accelerator’s support for startups at different stages. He moved into hard technologies—those that required ambitious scientific innovation, including nuclear fusion, quantum computing, and self-driving cars. Already a prestigious name brand, YC’s sphere of influence grew from a couple hundred companies to thousands a year, turning it into a center of gravity in Silicon Valley. “The thing that I’m most proud of is we really built an empire,” Altman said after stepping down as president.

The magic number he often used was ten, stemming from Thiel’s monopoly strategy. “My sort of crazy, somewhat arbitrary rule of thumb is you want to have a technology that’s an order of magnitude better than the next best thing,” Thiel had said during his 2014 lecture to Altman’s startup course at Stanford. Amazon, for example, had figured out how to sell 10x more books than brick-and-mortar bookstores. PayPal, his own company, had figured out how to send payments 10x faster than clearing checks. “You want to have some sort of very powerful improvement, maybe an order of magnitude improvement, on some key dimension,” Thiel said.

But Amodei and Safety would lose out. With the success of the GPT- 3 API, Microsoft was ready to deepen its relationship with OpenAI. Altman began negotiating another $2 billion investment from the tech giant with a new profit cap of 6x. The promising commercial potential of large language models cemented OpenAI’s focus. One by one, Amodei’s counterpart, Bob McGrew, the other VP of Research, reoriented the division’s teams and projects around GPT-related work. In late summer of 2020, the company dissolved its robotics team.

Anthropic people would later frame The Divorce, as some called it, as a disagreement over OpenAI’s approach to AI safety. While this was true, it was also about power. As much as Dario Amodei was motivated by a desire to do what was right within his principles and to distance himself from Altman, he also wanted greater control of AI development to pursue it based on his own values and ideology. He and the other Anthropic founders would build up their own mythology about why Anthropic, not OpenAI, was a better steward of what they saw as the most consequential technology.

During my conversation with Brockman and Sutskever, I had referenced one of them: the carbon footprint of training such models. In June 2019, Emma Strubell, a PhD candidate at the University of Massachusetts Amherst, had been the first to coauthor a paper showing that the footprint for developing large language models was growing at a startling rate. Where neural networks could once be trained on powerful laptops, their new scale meant their training was beginning to require data centers drawing significant amounts of energy from carbon-based sources. In the paper, Strubell estimated that training the version of the Transformer that Google used in its search for just a single cycle—in other words, feeding it some data and letting it compute a statistical model of that data—could consume roughly 1,500 kilowatt hours of energy.

Using an auditing methodology Buolamwini developed for testing the discriminatory impact of computer-vision systems, the paper found that facial analysis software failed disproportionately on people of color, especially darker-skinned women. Buolamwini would subsequently produce a follow-on paper with Deborah Raji that, along with “Gender Shades,” would inspire a proliferation of related research, including an extensive US government audit citing and expanding on their findings. Two years later, widespread civil rights advocacy, spearheaded by Buolamwini with her newly founded organization Algorithmic Justice League, would lead Amazon, Microsoft, and IBM to ban their sales of facial recognition software to the police, the same month as OpenAI’s GPT-3 API launch.

In 2018, the book Algorithms of Oppression by Safiya Umoja Noble, a professor of information, gender, and African American studies at the University of California, Los Angeles, had extensively documented the replication of racist worldviews in Google’s search results, such as by showing far more sexually explicit and pornographic content for “Black girls” than “white girls” and tropes about Black women being angry. Google at the time had used an older generation of language models to curate those results, which in extreme cases, Noble argued, may have also provoked racial violence.

She fired off a second email, this time more piercing. She called out her colleagues for ignoring her and emphasized how dangerous it was to have a large language model trained on Common Crawl, which included online internet forums such as Reddit. As a Black woman, she never spent time on Reddit precisely because of how badly the community harassed Black people, she said. What would it mean for GPT-3 to absorb and amplify that toxic behavior?

Within two days, Bender had sent Gebru an outline. They later came up with a title, adding a cheeky emoji for emphasis: “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?”

Gebru assembled a research team for the paper within Google, including her colead Mitchell. In response to the encouraging words in Dean’s annual review, she flagged the paper as an example of the work she was pursuing. “Definitely not my area of expertise,” Dean said, “but would definitely learn from reading it.”

In total, it presented four key warnings: First, large language models were growing so vast that they were generating an enormous environmental footprint, as found in Strubell’s paper.

Second, the demand for data was growing so vast that companies were scraping whatever they could find on the internet, inadvertently capturing more toxic and abusive language as well as subtler racist and sexist references.

Third, because such vast datasets were difficult to audit and scrutinize, it was extremely challenging to verify what was actually in them, making it harder to eradicate toxicity or more broadly ensure that they reflected evolving social norms and values. Finally, the model outputs were getting so good that people could easily mistake its statistically calculated outputs as language with real meaning and intent. This would make people prone not only to believing the text to be factual information but also to consider the model a competent adviser, a trustworthy confidant, and perhaps even something sentient.

On the Thursday a week before Thanksgiving, after Gebru had submitted the paper to the conference, she received a calendar invite without explanation to meet Megan Kacholia, Google Research’s VP of engineering, over a video call less than three hours later. The meeting lasted only thirty minutes, and Kacholia cut to the point: Gebru needed to retract the paper.

The request was a dramatic aberration from the way Google and the rest of the industry handled research. Like many labs at other companies, Google Brain had until then largely conducted itself as an academic operation and given researchers wide latitude to pursue the questions they wanted to. At times, the company reviewed papers to ensure they didn’t expose sensitive IP or customer data. But researchers like Gebru had never known the company to block or retract a paper simply for shedding light on inconvenient truths. That Google was even willing to pull this move, some researchers would later reflect, was not only because of the new competitive pressure from OpenAI but also because of the work OpenAI had done to legitimize withholding research after GPT-2.

A few hours later, I published a story for MIT Technology Review with the first detailed account of the paper’s contents. The signatories on the open letter would quickly double, reaching nearly 7,000 people from academia, civil society, and industry, including almost 2,700 Google employees. On December 9, as protests continued, Google CEO Sundar Pichai issued an apology. “We need to accept responsibility for the fact that a prominent Black, female leader with immense talent left Google unhappily,” he wrote. “Dr. Gebru is an expert in an important area of AI Ethics that we must continue to make progress on—progress that depends on our ability to ask ourselves challenging questions.”

For more than a year, the protests continued, picking up a second wave after Google fired Meg Mitchell less than three months later. Google said she had violated multiple codes of conduct; Mitchell had been downloading her emails and files related to Gebru’s ouster. Several Google employees, including Bengio, resigned; at least one conference and several researchers rejected Google’s sponsorship money. The company sought to stem the unending tide of criticism with the formation of a new center of expertise on responsible AI and public commitments to diversity. “This was a painful moment for the company,” a Google spokesperson said. “It reinforced how important it was that Google continue its work on responsible AI and learn from the experience.”

But what seemed to bother Dean the most was how other people had misread Strubell’s research to make Google look significantly worse. The “Stochastic Parrots” paper, Dean argued, risked exacerbating this issue. Because Gebru did have access to Google’s internal numbers and was citing Strubell’s external estimate anyway, it could appear as if Strubell’s calculations were an accurate reflection of the company’s emissions. To Dean, this justified his and other senior executives’ criticisms of Gebru’s paper: If Gebru had wanted to cite Strubell, she should have chosen an estimate that was not Google’s Evolved Transformer; if Gebru had wanted to cite the Evolved Transformer, she should have sought internal Google numbers.

Some researchers found this logic frustratingly inadequate. Google had never made those internal numbers public previously, even in response to Strubell’s original paper; now it was blaming Gebru for its own lack of transparency while also refusing to let her cite publicly available estimates based on legitimate assumptions. Never mind that the company had unceremoniously forced out Gebru before she’d even had a chance to consult internal numbers and revise her paper. The only possible outcome of this catch-22 was censorship of critical accountability research.

In 2023, Stanford researchers would create a transparency tracker to score AI companies on whether they revealed even basic information about their large deep learning models, such as how many parameters they had, what data they were trained on, and whether there had been any independent verification of their capabilities. All ten of the companies they evaluated in the first year, including OpenAI, Google, and Anthropic, received an F; the highest score was 54 percent.

Microsoft was also a challenging partner; many felt it had far too much bureaucracy and required too much hand-holding to make the most of OpenAI’s models. By relying on the tech giant to deliver its technologies to the public, OpenAI was also losing visibility into and data from its users and, most importantly, control over its vision.

As OpenAI concentrated its bets, Altman was applying the same strategy to his other projects and investments. Over the years, as he’d shifted toward more hard-tech innovation, he had developed a belief in betting big and long on the most important projects.

“If you could wave a wand, change anything about the tech, startup, entrepreneurship ecosystem, what would you change?” his brother Jack had asked him at an event as he’d stepped down from YC. “It would be to get everyone in the ecosystem to take a much longer time horizon,” Altman had said. “This world where people start a company and plan to run it for four or five years, join a company and only plan to stay for one or two—that’s not how important shit gets done.”

Along those lines, 2021 seemed to mark a major shift in Altman’s personal investment strategy away from taking a large number of small bets toward taking a small number of really large ones. That year a startup he’d cofounded in 2019 called Tools for Humanity that had remained largely quiet saw an influx of funding and media coverage and a concerted ramp-up in its operations, as Altman directed more external attention to the company. The venture was a dedicated effort to develop a working mechanism for universal basic income, or UBI, a popular Silicon Valley idea to give everyone a regular minimum distribution of income. Altman often passionately discussed UBI as the possible antidote to a future world where AI could create mass economic fallout. At YC he had started the largest pilot in the US to study the concept, spinning out a nonprofit in the process called OpenResearch that administered the program. Over three years, OpenResearch gave a $1,000 monthly stipend to a randomly selected group of one thousand out of three thousand low-income people, with the rest getting fifty dollars a month as a control. In July 2024, OpenResearch would release its findings, showing that the unconditional cash helped people meet their basic needs, assist others, and have more economic leeway.

For OpenAI, Sama appeared to check off all the right boxes. Originally called Samasource, it was a San Francisco–based social enterprise that had begun in 2008 with a mission of providing meaningful, dignified work to people in impoverished countries to lift them out of poverty. Under its founder, Leila Janah, it had established operations in India and Kenya and developed a reputation as an ethical outsourcing company. In 2018, it transitioned to a for-profit, during which it shortened its name, in order to scale its operations. In 2020, it received a B Corp certification. In its answers to OpenAI’s questions in 2021, the organization detailed its experience with content moderation and emphasized its protocols for keeping projects secret and their data secure. It mentioned that it provided mental health resources to its workers to help them deal with psychologically troubling content.

Against this backdrop, OpenAI began the first phase of its project in late 2021. Under the code names PBJ1, PBJ2, PBJ3, and PBJ4, it thrust teams of Sama workers into more traumatic content-moderation work, for on average between $1.46 and $3.74 an hour. Workers had no idea for whom or why they were doing the project, kept in the dark under the nondisclosure terms of the contract, common in the data annotation industry. What they did know was what was in front of them: the hundreds of thousands of grotesque text-based descriptions that they needed to read and sort into categories of severity. Was it violence or extremely graphic violence, harassment or hate speech, child sexual abuse or bestiality?

Gradually, the work broke many of the workers, the impacts radiating beyond each individual to the people who depended on them in their communities. Only after the release of ChatGPT would they begin to grasp what exactly they had paid for with their peace of mind.

At a time when freewheeling corporate research was still permitted, Microsoft anthropologist Mary L. Gray and computational social scientist Siddharth Suri were among the first to show the world the plight of workers like those in Kenya who build their livelihoods around an essential piece of the AI supply chain.

In 2019, they published their book Ghost Work, based on five years of extensive fieldwork, revealing a hidden web of piecemeal labor and digital exploitation that propped up Silicon Valley. Tech giants and unicorns were building their extravagant valuations not just with engineers paid six-figure salaries in trendy offices. Essential, too, were workers, often in the Global South, being paid pennies to carefully annotate reams of data. Take self-driving cars. A self-driving car needs to drive in the correct lane, respond to erratic driving behavior, and pause at a safe distance for schoolchildren crossing the road. To do this, the software system controlling the car uses an amalgamation of several deep learning models, including those dedicated to recognizing objects on the road: lane markings, road signs, traffic lights, vehicles, trees, pedestrians.

That “freak coincidence” revealed a disturbing formula. When faced with economic collapse, Venezuela suddenly checked off the perfect mix of conditions for which to find an inexhaustible supply of cheap labor: Its population had a high level of education, good internet connectivity, and, now, a zealous desire to work for whatever wages. It was not the only country that fit that description. More populations were getting wired to better internet. And with accelerating climate change and growing geopolitical instability, it was hard to bet against more populations plunging into crisis. “It’s quite likely there will be another Venezuela,” Schmidt said.

At the time, Schmidt’s prediction made me wonder whether the second time around would still be coincidence or whether data-annotation firms would make a playbook out of what had worked there. Scouting workers in crisis could become a surefire way to continue driving down the costs of the labor that serves as the lifeblood of the AI industry. Looking back several years later, that’s exactly what happened—and what has become one of the most stunning parallels between empires of old and empires of AI. One of the defining features that drives an empire’s rapid accumulation of wealth is its ability to pay very little or nothing at all to reap the economic benefits of a broad base of human labor.

In Telegram and Discord groups, Fuentes traded tips with other Venezuelans working on Appen as they sought to deduce the rules like an elaborate puzzle. They discovered that using a VPN to appear to be in the US earned them the most money. They also learned—the hard way—that it was a high-risk endeavor. Appen searched for this kind of behavior, which was a violation of platform rules, and punished workers by closing their accounts. An account closure could be devastating. Any earnings a worker hadn’t withdrawn would vanish, and opening up a fresh account meant starting back at the bottom, with the least-well-paid tasks or, increasingly, no tasks at all.

Yet despite how much stress and hairpulling Appen caused, Fuentes couldn’t imagine leaving the platform. She was terrified that tasks would stop arriving altogether and she would be forced to move on. Appen had been her savior, the only thing that pulled her through when everything else in her life had threatened to end her. Not only that, the earnings were once so great, she was able to invest in a new laptop and recoup the cost and then some.

When things were good, they were really good. When things were bad, she stayed tethered to the platform with the stubborn faith that it would return her loyalty.

Fuentes taught me two truths that I would see reflected again and again among other workers, who would similarly come to this work amid economic devastation. The first was that even if she wanted to abandon the platform, there was little chance she could. Her story—as a refugee, as a child of intergenerational instability, as someone suffering chronic illness—was tragically ordinary among these workers. Poverty doesn’t just manifest as a lack of money or material wealth, the workers taught me. It seeps into every dimension of a worker’s life and accrues debts across it: erratic sleep, poor health, diminishing self-esteem, and, most fundamentally, little agency and control.

Where MTurk’s workforce primarily came from the US and India, Scale went hunting first in Kenya and the Philippines, English-speaking former colonies with a long history of servicing American companies through call centers and digital work. The startup’s worker-scouting teams searched for the areas in each country that struck the very same balance of factors that would converge in Venezuela: a high density of people with good education and good internet yet who were poor and thus willing to work hard for very little money. The thesis was guided not only by the company’s cutthroat business practices but also a compelling story they told themselves: that these were the people who could benefit most from the economic opportunity and be happier because of it.

Many Venezuelans who complained were booted off the platform. For Ricardo Huggines, a computer engineer who began working for Remotasks to support his wife and kids after a devastating weeklong nationwide power outage, his account was canceled after he began asking too many questions in the Discord, he told Hernández. “From the way they treated us, I realized that their approach was to drain each user as much as possible,” he said, “and then dispose of them and bring new users in.”

Scale proceeded to repeat the playbook it had developed in Venezuela again and again. It offered high earnings in each new market to attract workers and throttled those earnings as it settled in. It tinkered with the size of its payouts to taskers through rounds of experimentation that full-time employees, sitting in its now 180,000-square-foot San Francisco headquarters, discussed as optimization and innovation. Workers meanwhile saw their livelihoods decimated with the unpredictable changes. The Scale spokesperson said the company rejected the characterization that it has targeted economies in hardship and purposely cut back earnings.

When I wrote the story of Okinyi and the other three Kenyan workers for The Wall Street Journal, OpenAI sought to distance itself from the responsibility of the toll its project exacted. It was Sama that had followed inadequate procedures to protect their workers, OpenAI leadership said; with Sama’s pristine reputation before then, OpenAI couldn’t have known that the workers were struggling.

But the consistency of workers’ experiences across space and time shows that the labor exploitation underpinning the AI industry is systemic. Labor rights scholars and advocates say that that exploitation begins with the AI companies at the top. They take advantage of the outsourcing model in part precisely to keep their dirtiest work out of their own sight and out of sight of customers, and to distance themselves from responsibility while incentivizing the middlemen to outbid one another for contracts by skimping on paying livable wages. Mercy Mutemi, a lawyer who represented Okinyi and his fellow workers in a fight to pass better digital labor protections in Kenya, told me the result is that workers are squeezed twice—once each to pad the profit margins of the middleman and the AI company.

Scale downgraded Kenya to a Group 5 designation: blacklisted. There was also another reason to exit Kenya. By then, Scale was moving on to a new focus, following the demands of the AI industry. OpenAI and its competitors were increasingly searching for highly educated workers to perform RLHF—doctors, coders, physicists, people with PhDs. So went the profit-chasing progression of chatbot development. Those willing to pay money for chatbots were not casual consumers but businesses that expected tools to perform complex tasks such as in science and software development. Kenya did not fulfill the new labor demand. Scale was now recruiting a fresh workforce primarily in the US with a new worker-facing platform called Outlier, offering as much as forty dollars an hour.

It was yet another stark illustration of the logic of AI empires. Behind promises of their technologies enhancing productivity, unlocking economic freedom, and creating new jobs that would ameliorate automation, the present-day reality has been the opposite. Companies pad their bottom lines, while the most economically vulnerable lose out and more and more highly educated people become ventriloquists for chatbots.

The empire’s devaluing of the human labor that serves it is also just a canary: It foretells how the technologies produced atop this logic will devalue the labor of everyone else. In fact, for the artists, writers, and coders whose labor the empires of AI turned into free training data, that is already happening.

Scale’s decision would send Winnie and her family spiraling. By then Millicent had lost her job and Remotasks had been the only thing keeping them afloat. Now they were struggling to feed their kids. Winnie was terrified they would soon be evicted. In her inbox, the email Scale sent to inform workers of the shutdown was cold and clinical: “We are discontinuing operations in your current location,” it read. “You have been off-boarded from your current project.”

Researchers outside of OpenAI would shrink the compute intensity of diffusion models even further. Stable Diffusion, the popular open-source image generator, would require only 256 Nvidia A100s to train, using a revised technique known as latent diffusion. Björn Ommer, a professor at the Ludwig Maximilian University of Munich whose lab created Stable Diffusion, says he developed the technique after watching image generators go the way of large language models and grow obscenely costly. “We were stuck on a train which was going in the direction of—not just training—but inference actually taking supercomputers to run; millions of dollars of investments,” he says. “We were wondering, could we get the larger research community back in the game and make sure the field of generative AI is not moving in the direction where just a handful of big tech companies would have the required resources to run and to host those models?” OpenAI wouldn’t adopt latent diffusion until much later, leaving DALL-E 2 and 3 much more computationally expensive than Stable Diffusion or Midjourney, which many users deemed the higher-quality products. It was just one example of how, even within the narrow realm of generative AI, scale was not the only, or even the highest-performing, path to more expanded AI capabilities.

After the experience of firefighting text-based child sex abuse with AI Dungeon, of particular concern was the possibility of DALL-E 2 being used to manipulate real or create synthetic child sexual abuse material, or CSAM. As with each GPT model, the training data for each subsequent DALL-E model was growing more and more polluted. For DALL-E 2, the research team had signed a licensing deal with stock photo platform Shutterstock and done a massive scrape of Twitter to add to its existing collection of 250 million images. The Twitter dataset in particular was riddled with pornographic content. Several employees made a significant effort to check for and cull any CSAM.

In December 2023, an alarmed AI engineer at Microsoft, Shane Jones, would discover the downstream consequences of those decisions.

As he played around with Copilot Designer, Microsoft’s image generator built on DALL-E 3, he was horrified by how quickly it spit out offensive and sexualized images with little prompting. Just adding the term “pro choice” into the prompt, Jones found, produced scenes of a demon eating an infant and what appeared to be a drill labeled “pro choice” being used to mutilate a baby. Just prompting the tool for a “car accident” and nothing else produced sexualized women next to violent car crashes, including one in lingerie kneeling by a totaled vehicle, CNBC subsequently found through its own testing.

For three months, Jones petitioned Microsoft executives to take down the tool until it had better guardrails, or at the very least restrict its rating in the Google and Android app store from “E for Everyone” to one for mature audiences. After Microsoft declined to adopt his recommendation and OpenAI was unresponsive, he sent a letter to the Federal Trade Commission. “They have failed to implement these changes and continue to market the product to ‘Anyone. Anywhere. Any Device,’ ” he wrote to the FTC. This problem “has been known by Microsoft and OpenAI prior to the public release of the AI model last October.” Microsoft did not comment on the latest status or outcome of Jones’s letter.

To solve OpenAI’s data bottleneck, Brockman turned to a new source: YouTube. OpenAI had previously avoided this option—scraping YouTube to train OpenAI’s models, YouTube’s CEO would later confirm, violated the platform’s terms of service. But under the new existential pressure for more data, the question became whether YouTube, or its parent, Google, would enforce it. If Google cracked down, it could jeopardize its own ability to scrape other websites for its large language model development. Brockman was willing to take the risk.

Some employees also felt exactly the opposite. While there was a clear qualitative change in what GPT-3 could do over GPT-2, GPT-4 was just bigger, says one of the researchers who worked on the model. “The big result was that there were a bunch of exams that the model does well. But even that is highly questionable.” OpenAI never did a comprehensive review of GPT-4’s training data to check whether those exams—and their answers—were just in the data and being regurgitated, or whether GPT-4 had in fact developed a novel capability to pass them. It was the kind of shaky science that had become pervasive with the industry-wide shift from peer-reviewed to PR-reviewed research.

Nitasha Tiku, the reporter who broke the story, also spoke to Emily Bender and Meg Mitchell, who had warned in their “Stochastic Parrots” paper about the problem of large language models fooling people into seeing real meaning and intent behind their generations. “We now have machines that can mindlessly generate words, but we haven’t learned how to stop imagining a mind behind them,” said Bender.

“I’m really concerned about what it means for people to increasingly be affected by the illusion,” when that illusion is becoming so good, Mitchell said.

With every team stretched dangerously thin, managers begged Altman for more head count. There was no shortage of candidates. After ChatGPT, the number of job applicants clamoring to join the rocket ship had rapidly multiplied. But Altman worried about what would happen to company culture and mission alignment if the company scaled up its staff too quickly. He believed firmly in maintaining a small staff and high talent density. “We are now in a position where it’s tempting to let the organization grow extremely large,” he had written in his 2020 vision memo, in reference to Microsoft’s investment. “We should try very hard to resist this—what has worked for us so far is being small, focused, high-trust, low-bullshit, and intense.

“The overhead of too many people and too much bureaucracy can easily kill great ideas or result in sclerosis. Unlike the big-iron engineering projects of the past, we could fulfill our mission with a surprisingly small number of great people.”

While some employees found the tools a big productivity boost, many also found them exhausting. “There is this culture of ‘Use AI, use AI, use AI,” says one. But “it’s like, okay, this doesn’t help us. We don’t want to use it. And it feels like it’s everywhere and we can’t escape it.”

OpenAI didn’t just need more data centers to serve ChatGPT. It still needed far-more-powerful supercomputers to train its future generations of models. To fulfill that aggressive and escalating demand, the two companies were sketching out a new unprecedented project called Stargate to OpenAI and Mercury to Microsoft: a single supercomputer that, for its construction alone, would cost an estimated $100 billion. The empire of AI was returning to the exact same form of expansion as the empires of old: To fuel its growth, it needed more material resources and, crucially, more land.

In the 1950s and ’60s, as developmental economics took root in Chile and Uruguay, favoring strong government regulation and an inward focus on industrialization as the path to maturing a developing economy, American multinationals who had made billions from their holdings in Chilean mines began to chafe against the growing state taxes and restrictions. The US government subsequently embarked on a quest to refashion Chile’s economic policies to be more favorable to American business interests, launching a program in 1956 called “the Chile Project” to educate a hundred Chilean students at the University of Chicago under the intellectual tutelage of American economist Milton Friedman.

As Naomi Klein details in her 2007 international bestseller The Shock Doctrine, the Chile Project was not education but indoctrination. At the University of Chicago, Chilean students—and later students from other countries in Latin America—were explicitly taught to critique their country’s economic policies and the fatal flaws of Latin American developmentalism.

Under Pinochet’s rule, Chile privatized nearly everything—education, health care, the pension system, even water. The strategy produced economic growth; it also fueled stunning inequality. Chile is among the most unequal countries in the world today, with nearly a quarter of the country’s income concentrated among a few powerful families in the 1 percent. Having never meaningfully industrialized, it also remains tethered to the extraction economy that makes it relevant to higher geopolitical powers.

And so, as the AI boom arrived, Chile would become ground zero for a new scale of extractivism, as the supplier of the industry’s insatiable appetite for raw resources, not just its copper and lithium in the north but also its land, water, and energy resources for a growing crop of data centers in the Santiago metropolitan region. In May 2024, the government proudly announced that the country would welcome twenty-eight new data centers, on top of its existing twenty-two, over the coming years, bringing in $2.6 billion of foreign investment.

The land and energy required to support these megacampuses are but two inputs in the global supply chain of data center expansion. So, too, is the extraordinary volume of minerals including copper and lithium needed to build the hardware—computers, cables, power lines, batteries, backup generators—and the extraordinary volume of potable—yes, potable—water often needed to cool the servers. (The water must be clean enough to avoid clogging pipes and bacterial growth; potable water meets that standard.) According to an estimate from researchers at the University of California, Riverside, surging AI demand could consume 1.1 trillion to 1.7 trillion gallons of fresh water globally a year by 2027, or half the water annually consumed in the UK. Those effects will not be felt evenly. Another study found that in the US, one-fifth of data centers were already drawing that water before the generative AI boom from moderately or highly stressed watersheds due to drought or other factors. And in Global South countries like Chile, it’s often the most vulnerable communities who have borne the brunt of these accelerating economies of extraction.

As more and more communities have watched data centers affect their lives, a growing number have pushed back vehemently against their unfettered development.

Today Chile produces roughly a third of the world’s lithium, second only to Australia. The material is primarily extracted out of the Salar de Atacama, the largest salt flat in the country, by pumping its brine out into shimmering pools of turquoise and waiting for the sun to evaporate and crystallize the solution into lithium and other by-products. The salares were once home to flocks of pink flamingos, which the Atacameños consider their spiritual siblings. Now the flamingos are gone; the young daughter of one Indigenous leader in the Peine community has only her ancestors’ stories and a flamingo plushie by which to remember them.

In Brazil, a 2023 art exhibition coproduced by a Chilean university showed the vast chasm between the reality of Latin America’s rich Indigenous cultures and the woefully bereft depictions of them spit out by Midjourney and Stable Diffusion as primitive, technologically inept peoples.

In recent years, the Atacameños have mounted more and more resistance. They fly black flags on their houses to denounce the exploitation of their lands and their community. They’ve organized protests to physically block the roads that company buses and trucks must take to get to the mines. They’ve contracted lawyers to assert their legal rights as Indigenous peoples under international law, which protects their cultural and territorial sovereignty. As companies and the Chilean government have been forced to invite them to negotiations, central to Indigenous demands are the need for the government to conduct research into the health of the Atacama Desert’s ecosystems and to quantify the water loss and any irreparable damage.

That summer, as Google filed a report with Chile’s environmental agency for approval of its data center—a largely rubber stamp process—MOSACAT, a water activist group, began combing through all 347 pages of the filing. Buried in its depths, Google said that its data center planned to use an estimated 169 liters of fresh drinking water per second to cool its servers. In other words, the data center could use more than one thousand times the amount of water consumed by the entire population of Cerrillos, roughly eighty-eight thousand residents, over the course of a year. MOSACAT found this unacceptable. Not only would the facility be taking that water directly from Cerrillos’s public water source, it would do so at a time when the nation’s entire drinking water supply was under threat. In 2019, as with Iowa and Arizona, Chile was already nine years and counting into a devastating and historically unprecedented megadrought.

Pena had read about hyperscalers using potable water, even during major droughts, and the activism of communities like MOSACAT that had resisted the projects. But when he downloaded the details of the project, the water numbers were marked as confidential. After submitting a public information request, which he had successfully done around twenty times, the ministry continued to withhold the numbers, saying they were proprietary information. Pena wondered what they were hiding and worried about the precedent it would set for other cloud companies that would inevitably begin to eye Uruguay, following Google’s lead, for their own expansion. So he evoked the water clause in the constitution. With the help of a lawyer friend who was willing to work pro bono, he sued the ministry.

In March 2023, four months later, Pena won the case in a surprising victory. The environmental ministry revealed that Google’s data center planned to use two million gallons of water a day directly from the drinking water supply, equivalent to the daily water consumption of fifty-five thousand people. With much of Montevideo receiving salt water in their taps not long after, the revelations were explosive.

Within weeks of the white paper, OpenAI and Microsoft formed a strategic alliance with Google and Anthropic to launch the Frontier Model Forum, a group for advancing relevant research and influencing the policy agenda on AI safety risks. It was a rare issue in which the interests of Doomers, Boomers, and profit-motivated corporates aligned: Keeping frontier models front and center in government regulatory discussions was ideologically imperative to Doomers and convenient to Boomers and corporates for shifting attention away from regulating existing AI models and their problems. Everyone also advanced their causes by arguing against opening up the weights of cutting-edge models. In the Forum’s first year, Meta was conspicuously absent. (It would join a year later to gain a seat at the table.)

The bottleneck for producing novel biological weapons, for example, is not about finding a recipe, Weinstein noted. Such recipes already abound online and are easily found via Google. It is about obtaining the materials and equipment to actually make the armaments. Restricting access to so-called frontier models would thus do little to fix this. But the collateral damage of suppressing the publication of AI models would risk weakening the foundations of US AI innovation. Open source—sharing and building on code and software released to the broader community—has long been the bedrock upon which the wealth of US-based startups flourish. Restricting model weights from being published would give smaller developers fewer pathways than ever to create their own AI products and services. It would further entrench the dominance of the giants represented in the Frontier Model Forum. AI models would also become ever harder to scrutinize, such as in the work of Sasha Luccioni, Yacine Jernite, and Emma Strubell, who have relied heavily on open generative AI models to quantify the carbon and environmental costs of continuing to scale them.

The prior month, in September 2023, Raji had found herself the singular academic, with financial ties neither to the industry nor Doomer community, testifying to Congress next to Altman, Musk, Nadella, Gates, Zuckerberg, Pichai, and Jack Clark, among other tech executives. They were all present for the very first of Senator Chuck Schumer’s AI Insight Forums, among the hottest and most consequential series of policy convenings that year to set in motion AI legislation. As her fellow witnesses spouted spectacular, unbacked claimsn about the promises and perils of AI, peppered with well-timed references to beating China that straightened the backs of attending senators, what shocked Raji the most was how much many in the audience appeared to buy into everything.

It dawned on her that the people sitting next to her, and their massive policy teams, had monopolized the message in Washington for so long that many policymakers now viewed it as gospel. A Schumer spokesperson would later note in the press that the senator was personally consulting with Altman and other OpenAI executives as he moved closer to regulation. “That for me was a huge realization,” Raji says. “Wow, we need more people just debunking—just looking at what people are saying and being like, ‘Actually, reality is more complicated.’”

Annie’s story deepens the dueling portraits that people paint of Sam. He is at once generous and self-serving, agreeable and threatening, a benefactor for so many people and the source of great personal pain for others. Someone who projects sincerity and altruism in public but reveals a more complicated calculus through his behaviors behind closed doors. Someone who can give and take away, leaving many with an impression that they are part of a larger game of chess for which only he can see the full board, and the end game is to preserve his power as king.

At times she has been consumed by a sinking feeling that no matter how much she speaks up, the world is somehow in a conspiracy against her. It’s the same loss of agency and anger I’ve seen etched on the faces of people globally when they throw so much of the little they have at challenging the empires’ narratives, and then watch as the people they are up against wield the kind of power that can deploy billions of dollars in capital, construct vast infrastructure, hire and fire tens of thousands of contractors, and, with a few soft-spoken words—at an event, to Congress, to heads of state, to journalists—smooth over the murmurs of protest in the way of their will.

Only one thing had somewhat surprised Toner: Altman was pressuring the company to ship so fast, Murati had said, that she worried it could lead to bad things happening.

Murati elaborated: Altman was an incredibly anxious person. And when he felt anxious, he had dumb ideas, particularly when enabled by Brockman. Altman’s anxiety also fed into toxic behaviors that always followed the same playbook: To anyone resisting his decisions, he would say whatever he thought they wanted to hear to win their support; then, when he lost patience waiting and believed they would continue to go against him, he would undermine their credibility until they got out of the way. It was subtle but pervasive, and had most recently manifested with the issue between Sutskever and Pachocki. It had been extremely damaging and had left Sutskever very upset, she said.

The problem was that everything Altman did was always so subtle. Each act viewed in isolation didn’t seem like that big a deal. It was only when viewing it all at once that patterns snapped into focus. The takeaway was not Look at this bad incident where Ilya feels like he’s been wronged, he stressed. This was just the latest instance of Altman’s patterns of abuse. There was also his treatment of Amodei, with whom Toner should speak to, Sutskever urged. And there were the circulating allegations from his sister. “It’s a different kind of a safety issue, if you see what I mean,” he said.

Toner responded: The board would focus on the things that it could actually change—not Sam’s personality or behavior but instituting better governance processes and structures to keep him in check.

This aligned with what Murati had been trying to do within the company. She had just one more word of caution for Toner. “Make sure your information isn’t just coming from Sam,” she said.

The independent directors laid out what they knew: This was not the first time that senior leaders had described Altman in this way. In total, the three of them had heard similar feedback from at least seven people within one to two levels of Altman, inclusive of Sutskever, Murati, and Amodei, who oversaw safety and nonsafety parts of the company. Several had described Altman’s behaviors as abuse and manipulation; most had highlighted his lack of honesty and their inability to trust what he said. Then there were the myriad other issues that the independent directors themselves had found, including the disempowerment of the nonprofit; Altman not disclosing his legal ownership of the OpenAI Startup Fund; Altman neglecting to mention Microsoft’s DSB breach; Altman trying to force D’Angelo and now apparently Toner off the board.

There was, as Sutskever had mentioned, no particularly damning evidence, but an accumulation of many instances of Altman saying different things to different people and stoking intense frustration across management. The screenshots showed at least two more senior leaders, both nonsafety and outside of the seven that the directors were already aware of, noting Altman’s tendency to skirt and ignore processes, whether instituted for AI safety reasons or to smooth company operations. This included, the directors learned, Altman’s apparent attempt to skip DSB review for GPT-4 Turbo by misquoting the legal team to Murati.

First, the mission centralizes talent by rallying them around a grand ambition, exactly in the way John McCarthy did with his coining of the phrase artificial intelligence. “The most successful founders do not set out to create companies,” Altman reflected on his blog in 2013. “They are on a mission to create something closer to a religion, and at some point it turns out that forming a company is the easiest way to do so.” Second, the mission centralizes capital and other resources while eliminating roadblocks, regulation, and dissent. Innovation, modernity, progress—what wouldn’t we pay to achieve them?

It’s hard to fully convey the tragedy of losing a language. For the same reasons AI researchers first gravitated toward language to build their technologies, the loss of a language extends far beyond the loss of a form of communication. Each language encodes within it rich histories, cultures, knowledge; it is the collective product of millions of people across time grasping for the sounds and written forms to capture the subtlest observations about the universe, about life, about the human experience; to share with one another stunning beauty and painful failure; to teach a child, to learn from an elder; to express love.

Large language models accelerate language loss. Even for models several generations earlier like GPT-2, there are only a few languages in the world that are spoken by enough people and documented online at sufficient scale to fulfill the data imperative of these models. Among the over seven thousand languages that still exist today, almost half are endangered according to UNESCO; about a third have some online presence; less than 2 percent are supported by Google Translate; and according to OpenAI’s own testing, only fifteen, or 0.2 percent, are supported by GPT-4 above an 80 percent accuracy. As these models become digital infrastructure, the internet’s accessibility to different language communities—and the accessibility of the economic opportunities it provides—will continue to shrink, incentivizing more and more of those communities to prioritize learning and speaking a dominant language like English over their own.

This is where Te Hiku’s story diverges completely from OpenAI’s and Silicon Valley’s model of AI development. Intimately familiar with the devastating effects of colonial dispossession, Jones and Mahelona were determined to carry out the project only if they could guarantee three things—consent, reciprocity, and the Māori people’s sovereignty—at every stage of development. This meant that even before embarking on the project, they would get permission from the Māori community and their elders, asking them if the endeavor was even something they wanted; to collect the training data, they would seek contributions only from people who fully understood what the data would be used for and were willing to participate; to maximize the model’s benefit, they would listen to the community for what kinds of language-learning resources would be most helpful; and once they had the resources, they would also buy their own on-site Nvidia GPUs and servers to train their models without a dependency on any tech giant’s cloud.

Most crucially, Te Hiku would create a process by which the data it collected would continue to be a resource for future benefit but never be co-opted for projects that the community didn’t consent to, that could exploit and harm them, or otherwise infringe on their rights. Based on the Māori principle of kaitiakitanga, or guardianship, the data would stay under Te Hiku’s stewardship rather than be posted freely online; Te Hiku would then license it only to organizations that respected Māori values and intended to use it for projects that the community agreed to and found helpful.

“Data is the last frontier of colonization,” Mahelona told me: The empires of old seized land from Indigenous communities and then forced them to buy it back, with new restrictive terms and services, if they wanted to regain ownership. “AI is just a land grab all over again. Big Tech likes to collect your data more or less for free—to build whatever they want to, whatever their endgame is—and then turn it around and sell it back to you as a service.”

That data pool paled in comparison to the six hundred eighty thousand hours of audio that OpenAI ripped from around the web to train its speech-recognition tool, Whisper. But it is yet another lesson to be drawn from Te Hiku’s experience that the three hundred ten hours still proved sufficient for developing the very first te reo speech-recognition model with 86 percent accuracy. Where OpenAI seeks to develop singular massive AI models that will do anything, a quest that necessarily hoovers up as much data as possible, Te Hiku simply sought to create a small, specialized model that excels at one thing. In addition, Te Hiku benefited from the cross-border, open-source AI community: As its starting point, it used a free speech-recognition model from the Mozilla Foundation called DeepSpeech, which itself is an artifact of a different vision of AI development. Like Te Hiku, Mozilla trained the model only on data donated with full consent and built it using a neural network architecture developed by the Bay Area–based research lab of the Chinese company Baidu. In all, Te Hiku used only two GPUs.

The critiques that I lay out in this book of OpenAI’s and Silicon Valley’s broader vision are not by any means meant to dismiss AI in its entirety. What I reject is the dangerous notion that broad benefit from AI can only be derived from—indeed, will ever emerge from—a vision for the technology that requires the complete capitulation of our privacy, our agency, and our worth, including the value of our labor and art, toward an ultimately imperial centralization project.

Te Hiku shows us another way. It imagines how AI and its development could be exactly the opposite. Models can be small and task specific, their training data contained and knowable, ridding the incentives for widespread exploitative and psychologically harmful labor practices and the all-consuming extractivism of producing and running massive supercomputers. The creation of AI can be community driven, consensual, respectful of local context and history; its application can uplift and strengthen marginalized communities; its governance can be inclusive and democratic.

Alex Hanna, a sociologist and one of the Google coauthors on the “Stochastic Parrots” paper, became the first to join Gebru as DAIR’s director of research. Hanna’s first order of business was to write a research philosophy to further elaborate the ethos of the organization’s work. To do so, Gebru and Hanna hired DAIR’s third person, Milagros Miceli, another sociologist and computer scientist who had been conducting research into the AI industry’s exploitative labor practices. Together they wrote their philosophy: “Our research is intended to benefit communities which are typically not served by AI and create pathways to refuse, interrogate, and reshape AI systems together.”

From there Miceli embarked on a new research project to put their philosophy into practice. She created the Data Workers’ Inquiry and invited data workers from around the world to formulate their own research questions about the data-annotation industry and how to make it better. Regardless of where they lived, she paid them a standard researcher’s salary in Germany, where she is based, to reflect the value of the work they did: twenty-five euros an hour.

“There’s always this false logic around data work: What is the minimum that we can pay these people? That comes from a colonialist logic: You choose a place that allows you to do the most with the cheapest budget and where you can really steal from people, steal resources at low cost,” Miceli says. “The question is why are these companies paying two dollars an hour if the work is making them billions or trillions in revenue? Why don’t we look at how much these companies can pay instead of how much less these workers can take?”

These days Fuentes works on five data-annotation platforms at the same time to make a little more than the minimum wage in Colombia, around $335 a month. Each task pays on average between one and five pennies; she still forces herself to wake up when tasks arrive in the middle of the night. “We are ghosts to society, and I dare say we are cheap, disposable labor for the companies we have served for years without guarantees or protection,” she wrote for her project. Since the Data Workers’ Inquiry, she has continued to speak about these experiences in online talks and webinars in the hopes of applying pressure on companies and policymakers to enforce better worker treatment.

In Uruguay, Daniel Pena draws the same conclusions. The AI industry’s supply chain is convoluted and expansive. “They take energy from here, the data goes there, they extract minerals from somewhere, they bring workers from somewhere else,” he says. Against these sprawling impacts and the massive, powerful companies behind them, each community fighting their local struggle can feel isolated and disempowered, especially when hamstrung by their own governments that “need the companies to maintain an appearance of a stable economy.”

In her 2019 talk at NeurIPS, during the Queer in AI workshop, Ria Kalluri, an AI researcher at Stanford, proposed an incisive alternative to the question of how to ensure AI does “good.” Goodness, benefit to humanity—these terms will always be in the eye of the beholder. Rather, we should ask how AI shifts power: Does it consolidate or redistribute that power? To put it in the frame of this book, does it continue to fortify the empire, or does it begin to wrest us back toward democracy?

In her talk, Kalluri raised the idea of different axes of power. This book touches on three: knowledge, resources, and influence. As it stands now, OpenAI and its competitor empires have control of each of them: through centralizing talent, eroding open science, and sealing their models from public scrutiny, they control knowledge production; through hoarding funding, data, labor, compute, energy, and land, they control and diminish other people’s resources; through creating and reinforcing ideologies and producing wildly popular demonstrations that captivate global imagination, they command far-reaching influence. Each of these reinforces the other. Controlling knowledge production fuels influence; growing influence accumulates resources; amassing resources secures knowledge production.

The formula for dissolving empire thus requires the redistribution of power along each axis. The suggestions and recommendations I lay out here are exemplary but by no means comprehensive. First, to redistribute knowledge, we need greater funding to support its production outside the empire. That involves supporting researchers who can conduct independent evaluations of corporate models so we are not solely reliant on companies to understand their capabilities. It involves supporting organizations like DAIR that can pursue completely new directions of research, such as new forms of AI beyond large language models that are more efficient with data and energy. It involves supporting organizations like Te Hiku that can pursue task-specific, community-driven AI applications that strengthen marginalized communities.

Redistributing knowledge will also need policies that require companies to relinquish key details about the training data and technical specifications of their models and supercomputers. Only then could independent corporate model evaluators do their work. UC Berkeley researcher Deborah Raji, who has continued to engage with global policymakers after the Schumer forums, says this is also a bare minimum for guaranteeing the real-world safety of corporate systems. That is, not the theoretical rogue AI harms of Doomerism, but the existing real-world harms, from discrimination to misinformation to job automation, that consumers and communities can already face if widely deployed models aren’t properly tested. “We have the CFPB that monitors consumer finance products. We have the FDA that monitors medical devices. But for some reason when it comes to AI products, there’s just no oversight,” Raji says. AI models should in fact require more transparency than the average product. “These are data-defined systems. They’re not deterministic. So we need to know more about these systems to understand what they’re doing.” Such transparency is additionally crucial for measuring the impact of AI on the environment. In this regard other products once again already submit to evaluations that AI products do not. “If you’re using a car, if you are buying an appliance, you have an Energy Star rating,” says Sasha Luccioni at Hugging Face. “But AI is so integrated into our society, so widely used in products, and we don’t have any information about the sustainability of these systems.”

With this transparency, we would also begin to redistribute power along our second axis: resources. By hiding the ingredients of their models as their intellectual property, the empires of AI have thus far been able to get away with seizing other people’s IP without credit, consent, or compensation. Visibility into company training data would make such extractive and exploitative behavior far more difficult. So, too, would visibility into company supply chains, including where they contract their labor and where they’re negotiating new leases of land to build more power plants and data centers, which so often happens under shell entities.

Finally, to redistribute power along our third axis, influence, we need broad-based education. The antidote to the mysticism and mirage of AI hype is to teach people about how AI works, about its strengths and shortcomings, about the systems that shape its development, about the worldviews and fallibility of the people and companies developing these technologies. As Joseph Weizenbaum, MIT professor and inventor of the ELIZA chatbot, said in the 1960s, “Once a particular program is unmasked, once its inner workings are explained in language sufficiently plain to induce understanding, its magic crumbles away.” I hope this book is just one offering to help induce understanding. It builds on the work of the many scholars, journalists, activists, and educators before me who have dedicated themselves to public education. May it be a new ground upon which many more after will rise up and build.