The internet could be on the verge of a total meltdown. Yesterday, someone posted to a public forum that two major security flaws had been found in a programming language that undergirds the web. The language, called PHP, is used by as much as 70 percent of websites.

In order to protect themselves from hackers, a number of government, hospital, and school websites across Europe went dark this morning. The organization that maintains PHP quickly confirmed it was preparing to announce two major vulnerabilities alongside patches for them. Both weaknesses, officially disclosed today, appear to have existed in PHP for more than a decade but were only just discovered with the use of AI tools. Hackers have already been detected trying to exploit them, likely with the help of AI, and implementing the fixes to address the vulnerabilities could take companies days or weeks. In the meantime, tens of millions of servers, databases, and devices are exposed. “PHP is one of the fundamental pieces of technology that supports the internet,” Sam Rubin, a senior vice president at the cybersecurity firm Palo Alto Networks, told me. As of now, the websites of the Finnish health agency, the Port of Rotterdam, the Paris police force, 4Chan, and an unknown number of others are all down while they address the PHP vulnerabilities.

[Read: How long until AI hacks everything?]

News of these weaknesses in PHP is the capstone in a weeks-long slew of devastating, AI-powered cybersecurity events. In recent weeks, criminals and state-backed attackers have used AI to hack into more than 100 companies, including John Deere and SharkNinja; to automate hacks on some 400 financial institutions in Latin America; and to steal hundreds of thousands of credit-card records from two companies. Since this summer, attackers have hit governments, data centers, and hospitals in Ukraine, Russia, Argentina, Bangladesh, and Indonesia. Also today, the developers of Node—a platform used by nearly half of software developers—disclosed multiple high-severity vulnerabilities. And that’s the abridged list.

For the better part of a year, cybersecurity experts have been warning, with mounting alarm, about an impending cyber-apocalypse. It may have begun. For decades, the internet has been riddled with holes and bugs. Code is finicky, and the standard IT approach has been to fix problems when they arise. For decades, the balance held: Skilled hackers are few, and finding and exploiting vulnerabilities takes time.

Recent advances in AI have totally upended that equilibrium. Top models have become very good at finding holes and writing malware, and can work much faster than even a team of humans. This is true of both open-source AI systems from China, which can be downloaded and used without restrictions, and the even more capable models from OpenAI and Anthropic, which are continually being jailbroken. The rate of software vulnerabilities identified per month has doubled over the past year. Meanwhile, the time that hackers need to begin exploiting those vulnerabilities—figuring out how to break in and out without being caught—has fallen, by some estimates, from weeks to hours. An actual intrusion, if automated with AI, can take as little as 39 seconds. Historical approaches to detecting and fixing vulnerabilities, or monitoring for suspicious internal activity, are insufficient.

AI has created new vulnerabilities, but it also promises to fix them: The vulnerabilities in PHP were discovered independently by a Google-owned cybersecurity company and another effort led by OpenAI. Google, OpenAI, and Anthropic, along with other tech firms, are all working with government agencies around the world to shore up their defenses. But, as numerous cybersecurity experts have told me, there is an asymmetry between attackers and defenders that AI is going to exacerbate: Hackers will use new technology with abandon, while companies are methodical and risk-averse—a hospital or electrical grid can’t risk that an AI security tool unintentionally takes a patient off life support or causes a brownout. And attackers need to find only one hole to succeed, whereas defenders need only one oversight to fail.

For a year, criminals and state-backed groups have been experimenting with how to use models to accelerate and automate their attacks. It was only a year ago that Anthropic announced, with great consternation, what it declared to be the first instance of AI-orchestrated cyber-espionage. That still involved a lot of human oversight and required jailbroken Claude agents to work for many hours. Now that the agents are faster and more capable, hackers have gotten more adept at deploying them, and the automation is more extensive. Recent, widespread attacks on South Korean and Japanese websites were reportedly executed by a single Chinese cybercriminal who used AI to target more than 1,000 websites. “This certainly will not be the last large-scale vulnerability identified, because of what’s possible with AI,” Rubin said.

The past 24 hours represent a tipping point. Bulking up IT departments, implementing passkeys, and using two-factor authentication remain important. But the most urgent and practical advice being given now is to retreat: Prepare to simply go offline. The NSA advised companies that cannot immediately fix the PHP vulnerability to take down affected digital services. Hospitals, water utilities, and other crucial services are all preparing for hacks by relearning to operate analog. Store your data in fewer places. We’re running out of time to prepare for the AI-hacking deluge. The dam may have already broken.

Up until a few years ago, you could be forgiven for never having heard much about data centers. When we connected to the many digital services we’ve come to depend on, and executives took to the stage to show off the latest improvements to their products, there wasn’t much talk of the physical infrastructure that made it all possible. But behind the Uber rides, the Amazon shopping, and the countless searches on Google, unremarkable buildings in industrial parks, stuffed full of hard drives and servers, were keeping it all  running.

It’s almost staggering to see how quickly data centers have become central to the national discourse. As the number and size of data centers exploded over the past few years, a once seemingly invisible infrastructure rapidly found itself at the center of a political and technological debate. The largest facilities, known as “hyperscale” data centers, can have hundreds of thousands of servers beneath a single roof, and they demand hundreds of megawatts of electricity from the grid—similar to a medium-size city. That scale brings a raft of challenges including pressure on electrical grids, high demands on the water system, and noise and light pollution for people living in their vicinity.

If only a few data centers operated, they might not have become such a significant political issue. But that’s the thing: Since the release of ChatGPT at the end of 2022, the tech industry has been on a building spree. In turn, the public has woken up to the presence of data centers—and people are not happy about them. A poll in August found that 75 percent of Americans opposed the development of any such projects near where they lived. In a short span of time, these once-overlooked facilities have become the core of a whole new political battleground. Data centers are set to be a key topic in the upcoming midterms, as well as the 2028 presidential elections.

In December of last year, Senator Bernie Sanders called for a national moratorium on data-center construction. “This process is moving very, very quickly,” he said in a video laying out his plan, “and we need to slow it down.” Sanders argued that data centers were part of a wider set of challenges that AI presented for workers’ livelihoods and for people’s personal relationships. In his view, AI was being driven by billionaires such as Jeff Bezos, Elon Musk, and Mark Zuckerberg, who were enhancing their power while openly pushing for technologies that could take people’s jobs and erode human connection, and that did not seem like a healthy path forward. “This moratorium will give democracy a chance to catch up with the transformative changes that we are witnessing,” Sanders said.

This wasn’t just a progressive issue. Donald Trump may regularly praise the AI industry and the data centers that power it, but that hasn’t stopped Republicans from opposing them too. The same month that Sanders proposed a moratorium, Florida Governor Ron DeSantis launched his own broadside against the AI industry. He wanted to empower local governments to reject data-center construction in the state and bring forward a set of standards to limit the risks caused by AI. He assured voters that his campaign had nothing to do with “woke” ideology but was about protecting workers from job loss, communities from data-center encroachment, and society from the “mindless slop” and falsified content generated with AI that could pose “a potential existential crisis for self‑government.”

By the time 2026 rolled around, soaring electricity prices had become a national concern as tech companies drove up energy costs in their pursuit of generative AI. Hundreds of local moratoriums were being passed across the country; in July, New York adopted the first statewide pause on hyperscale data-center projects. As public anger grew, communities were finding that they had some prominent allies. Such sentiments did not end at the U.S. border.

Restrictions on data-center development have been part of the conversation in certain parts of the world since the late 2010s. Singapore and the Dutch municipalities of Amsterdam and Haarlemmermeer both issued moratoriums on new projects in 2019, and by 2021, Ireland had introduced bans in certain areas because energy demands were outstripping supply. Some of these moratoriums were later lifted after stricter rules came into place, and over fear of missing out on the investment pouring into AI. Even still, these efforts have only grown since the data-center boom that was spurred by the dash to win the generative-AI race.

In recent weeks, Denmark downgraded data centers to the lowest priority for grid connections, and Thailand placed a pause on further data-center construction—suspending 49 projects where shovels had already broken ground. These moves demonstrate the tension that many governments are facing: trying to balance a desire to push the adoption of generative AI with the reality that the public is not on board with the massive data centers needed to power them.

In Europe, there’s also been a growing discussion about reducing dependence on U.S. tech companies. In the weeks after Sam Altman unveiled his Stargate plans at the White House, European governments shot back with billions of euros in planned data-center investment of their own. The social and environmental costs were not up for debate; politicians just wanted to keep up. But now many are being forced to reckon with the costs that come with this infrastructure.

Those fights are also playing out in the home of the major tech companies. For all the support that Trump gives the data-center industry—he has said that communities opposing such projects want to be “backwards and poor”—the reality is that politicians in states where the most data centers are clustered are being forced to change their positions as they feel the political winds shifting. Higher power bills and stories about what it’s like to live next to one of these hyperscale data centers have incensed voters. Meanwhile, the industry pours millions of dollars into electoral races to try to avert stricter regulations. The outcome of the November midterms will determine more than just what the next two years will hold for Trump; they could also decide the fate of the data-center boom across the U.S.

This article has been adapted from Paris Marx’s new book, Hyperscale: The Ambition and Excess of Big Tech’s Data Empires.

Ten years ago, I found myself in Lusaka, Zambia, sitting at a table with a group of extraordinary writers. We had been invited to participate in a 12-day writing retreat by the Caine Prize for African Writing, considered by many to be the most important literary award for authors on the continent. Among our ranks at the retreat were four former winners of the prize (NoViolet Bulawayo, Okwiri Oduor, Namwali Serpell, and me), as well as past finalists and immensely talented Zambian writers. During those two weeks, we wrote and workshopped our stories, and in between—as we ate and gossiped, visited local schools, and read our work to the public—we discussed the great distance between the literature we hoped to produce and the kind of work the marketplace expected from us.

We were all beneficiaries of the Caine Prize, which was first awarded in 2000 to introduce African writing to a wider international audience. At the time, African fiction hardly existed in Western bookstores or newspaper book sections. In the years since, it had flourished, but many of the most successful and celebrated works circled the same themes: war and its aftermath, the scourge of poverty, the persecution and murder of queer people, the trials of immigration to the West, recapitulations of popular American or European stories in African settings.

In Lusaka in 2016, we wondered whether there was room for stories that did not hew to these norms, stories that followed their own idiosyncratic paths, stories that challenged and reconceived what “Africa” means. So many of us, after all, were living examples of the diversity and capaciousness of the term. Even as we pursued originality in our work, we recognized that certain stories (I described them as “accessible” in an essay for the Los Angeles Review of Books) would continue to earn a disproportionate amount of recognition and acceptance.

A decade later, the Caine Prize has become the setting for a different conversation about African writers and literary authenticity. Two weeks ago, the Zimbabwean writer Simbarashe Steyn Kundizeza was named this year’s winner for his story “The Blue Fighter.” Just five days later, the novelist and journalist Vauhini Vara reported in these pages that the AI detector Pangram had flagged almost the entire story as AI-generated.

The allegation quickly became fodder for arguments far beyond the Caine Prize: about the capabilities of artificial intelligence, the state of contemporary literature, and even the legitimacy of Black intellectual achievement. Vara’s story revived fears raised earlier this year, when three regional winners for the Commonwealth Short Story Prize were accused of using AI, that chatbots are already winning prestigious literary prizes by fooling multiple award juries. Critics also worried that contemporary literature had grown so stale that even a machine was capable of creating award-winning prose. But the racial dimensions magnified the controversy. Some people contended that AI-detection tools were being used to unfairly undermine Black and brown authors. For others, the accusations confirmed a long-standing belief that Black and brown writers—and, by extension, Black and brown people—are incapable of achieving prominence without substantial assistance, either from institutions that supposedly favor them over more qualified individuals or, now, from artificial intelligence.

[Read: Yet another literary prizewinner seems to be AI]

Lost in these debates is a quieter, more unsettling irony: After years of being encouraged by the literary ecosystem—editors, prize committees, magazines, critics—to produce work that is accessible to white audiences, African writers are now facing a new kind of suspicion about their authorship. In short, we have arrived at a moment when writing that conforms to accepted literary conventions—including, in particular, those associated with “African” writing—can also seem suspiciously contrived.

Kundizeza’s story provides a signal example. I found it moving and quite effective in moments, but I also instantly recognized—and predicted—many elements of its plot. The story is narrated by a boy in Zimbabwe whose father abandons his career as a civil engineer to pursue his dream of building aquariums and breeding exotic fish. While his father pours the family’s dwindling resources into the tanks and their expensive equipment, the boy’s mother struggles to keep the household afloat.

Perhaps you can imagine what happens next: The family moves to a poorer neighborhood, the mother finally confronts the father about his obsession, and, over his objections, she sells nearly all of his remaining fish tanks to pay the bills. Kundizeza’s story had many components of prize-winning African literature: a youthful narrator (my own Caine Prize–winning story, “Miracle,” featured one as well); the struggling father with big, impossible dreams; the responsible and enterprising mother who holds everything together; the backdrop of mounting economic hardship.

Kundizeza has denied using AI and has said that he drew on characters and settings from his own life in writing the story. I don’t doubt this, but one reason the story might strike some as AI-generated is that it so neatly encapsulates and deploys the tropes we have come to expect from prize-winning African fiction. If chatbots do anything especially well, it is exactly this: ingesting our preferred tropes and delivering them right back to us.

Readers who raise suspicions about AI usually flag patterns in prose. But there is another pattern worth considering. Chatbots have learned from massive quantities of human writing, including representations of Africa shaped by the literary marketplace. It seems reasonable to expect that if you ask a chatbot to write an “African” story, it will reproduce familiar conventions that have been rewarded for decades.

[Read: A turning point in AI writing]

That said, I don’t believe that a Pangram result should be enough, by itself, to condemn a writer. Earlier this year, the Commonwealth Foundation investigated allegations against several of its regional winners, examining drafts, notes, and time-stamped documents before concluding that AI had not been used to write their stories. In a recent New York Times essay, the Cambridge professor Doyle Calhoun noted that the five literary controversies involving AI that he considered most significant this year all concerned Black and brown writers. The Kundizeza accusation adds another. Whatever the outcome of the case, these controversies reveal a troubling predicament: Even as the literary world struggles to determine what constitutes AI-assisted writing, Black and brown writers have repeatedly been forced to defend the authenticity of their work.

One of the problems with adjudicating these situations is that no one has spelled out what, exactly, constitutes cheating. It is virtually impossible to open up a Microsoft Word or Google document without encountering AI tools that encourage you to rewrite your prose. If you accept that assistance and your work is subsequently flagged by a detection program as AI-generated, have you cheated? What if you drop a completed story into a chatbot text box and request edits? Established writers have the benefit of publishing their work in well-resourced periodicals where it is edited by multiple humans; the resulting article, regardless of how many sentences were drafted, deleted, or polished to a sheen by other hands, is credited to one person. Is it wrong for an emerging writer to seek similar assistance from artificial intelligence? Many people are quick to accuse individual writers of cheating, but technology companies have integrated AI assistance into so many parts of our lives, and the literary world has yet to agree on where assistance ends and authorship begins.

My greatest worry is that the arrival of AI has imposed yet another set of expectations on Black and African writers. Not only must our work satisfy the literary marketplace, but it must also distinguish itself from the machines that have learned to imitate its conventions. Over the years, the establishment has rewarded predictability and accessibility—the very qualities that now make writing suspect. As our Zambian retreat drew to a close a decade ago, many of us reached a conclusion that seemed almost impractical then but feels urgent and essential now: We had to ignore the noise and market signals and, to borrow from the South African anti-apartheid activist Steve Biko, write whatever we liked.

It was fitting that, just hours after the first public screenings of the Elizabeth Holmes documentary began last Thursday, Elon Musk put on a show of his own.

That night, in Texas, Musk unveiled the company’s new Roadster prototype, a car that would defy gravity. Tesla built what was roughly one-half of a roller-coaster loop, which the Roadster climbed before momentarily hanging upside down and then executing a controlled flip and gliding down to Earth. The demo worked. In another stunt, it did a barrel roll over a semi-truck. Each maneuver was powered by air thrusters that guided the car to a soft landing.

The demo was impressive. But watching the clips, I could feel my brain struggling to process what I was being shown. Was this real? The car did climb vertically, flip, and land. But Musk, who has built wildly successful companies, is also famously prone to outlandish claims that don’t always pan out. I couldn’t help but wonder what, if anything, Tesla’s demonstration or Musk had elided. “This will be an actual car that’s on the roads,” Musk said. He got the flying-car headlines he wanted, but will this car actually be released? There are already reasons to doubt it: The car displayed on Thursday won’t be in production until at least 2028, and it features 10 thrusters powered by 600 liters of compressed air. The flying version of the Roadster will be sold “invite-only.” Its price tag is likely to be astronomical. And so Musk isn’t exactly lying here. Rather, he’s making bold proclamations and asking people to take his pledge to deliver as an article of faith.

[Read: ]Imagine every AI agent on Earth buying the same cheap paper towels

Less than 48 hours later, I found myself sitting in a crowded theater, watching the new documentary You Can See Everything, about the disgraced and convicted founder of Theranos, Elizabeth Holmes, and her husband, Billy Evans. In one scene, Holmes and Evans are hurriedly trying to court investors and research talent to build out what they refer to as Theranos 2.0. This scene is arguably the only time in the film where the duo don’t appear to be filled with existential dread. During phone calls and Zoom meetings, they spit out jargon and look earnestly excited about what they can build if they “get the band back together.” At the end of the sequence, Holmes is asked about why she’s spending her last days with her family building a company, and she responds that this is what she’s meant to be doing here on Earth. It is one of the few moments where I confidently believe her. Whatever she earnestly believes about her own guilt, it seems clear that her identity is inextricably tied to her notion that she is, in Silicon Valley parlance, a builder: a high-agency, visionary type with a remarkable risk tolerance who makes things and isn’t afraid to fail, because she has a monomaniacal obsession with pushing the bounds of what’s possible.

The film, which had the biggest opening for a documentary in more than 20 years, is extremely well suited to a post-truth era. The co-directors, Nathan Fielder and Lance Oppenheim, bury the audience with unreliable narration such that, when you get to the end, you feel like a participant in the film. It also left me exhilarated, slightly repulsed, and feeling as if I was going insane. In that way, the documentary replicates what it’s like to navigate the informational hall of mirrors we all inhabit. But what I found most bracing about the film—more so than the supposed unknowability of anyone or anything—is how Holmes’s act and affect do not feel like an aberration in 2026. In fact, her behavior embodies the current ethos of Silicon Valley.

When she built Theranos, Holmes was a college dropout with boundless ambitions who styled herself after Steve Jobs. Jobs famously said things such as “People don’t know what they want until you show it to them,” and Holmes, like Jobs, projected the aura of a person ready to pull the future into the present by any means necessary. Jobs routinely delivered on these promises, and kept many projects secret until they were ready for the masses. Holmes did the opposite, offering a fantastical version of a product that did not exist, obscuring the actual capabilities and lying about the financials. She wasn’t alone; over time, the idea of the builder became more synonymous with a founder’s ideas, charisma, and salesmanship than with the products themselves. “I watched countless founders of hard tech companies present fraudulent products at conferences,” Nikita Bier, a tech entrepreneur who was recently Elon Musk’s head of product at X, wrote on the platform this weekend in response to the documentary. “They had fake studies with fudged numbers. Most of the adults in the room excused it as ‘faking it ‘til you make it.’”

Holmes’s high-profile failure and the collapse of Theranos in 2018 didn’t kill the builder archetype. If anything, the very structure of Silicon Valley became more deeply intertwined with the builder mindset. Fake it ’til you make it (provided it isn’t outright fraud) is a necessary outcome of an industry that’s funded by venture capitalists, who often make strategic bets on projects so early in development that the products are still hypothetical. And so VCs threw unfathomable money at promising but half-baked ideas in the hopes of fabulous payouts. Failures were overlooked, even celebrated as necessary or teachable moments—honorable proof of being a “man in the arena.” In 2022, after Sam Bankman-Fried’s crypto empire was revealed to be an elaborate Ponzi scheme, the writer John Ganz argued that, in an industry built on big bets, the line between frauds and visionaries becomes difficult to suss out until everything implodes. “This is the Fyre Festival era of capitalism: outlandish promises and then a ludicrous debacle.”

Much of that rush to invest has been driven by a culture progressively gripped by the fear of missing out on the next industry-shaking payout. Untold billions were poured into crypto, then NFTs, Web3, and the ill-fated metaverse. In crypto, rug-pulling scams and Ponzi schemes abounded; its boosters said things such as “Have fun staying poor.” Today we’re living through what many suggest could be an AI bubble, and the same mentality persists. Have fun staying poor has been replaced with the idea of a “permanent underclass” of people who failed to get in early and now stand to lose their economic viability, and potentially their human agency altogether. Like Theranos, crypto, and NFTs, the AI industry is defined by claims of its inevitability and its future power (to be certain, chatbots and coding agents are real and powerful in ways the other technologies were not). In this way, AI is a faith-based technology; one’s desire to work on it, bet the success of the global economy on it, love it, or fear its End Times powers is typically predicated on what one can imagine it might become.

The past eight years have warped the tech industry to the point that it can feel untethered from the constraints of reality. You can see it in headlines everywhere if you’re looking. On Saturday, an Anthropic co-founder, Jack Clark, suggested that the U.S. GDP could jump in the next four years from about $30 trillion to $44 trillion thanks to rapid AI-driven progress in many industries. This would be one of the largest periods of growth in U.S. history, a prediction based on nothing more than the most “extreme parameter setting” of an Anthropic economic model the company released this fall. You can see it in the IPO prospectuses for a company like SpaceX, which includes in its ambitions the goal to “extend the light of consciousness to the stars” and initiatives such as asteroid mining and space tourism. And earlier this month, Musk noted that “this will sound super crazy, but I see a path to SpaceX being worth orders of magnitude more than the current Earth economy.” You could call all these claims tech optimism or, as the famed venture capitalist Marc Andreessen wrote in his 2023 manifesto, necessary for the “glory of human ambition and achievement, the spearhead of progress, and the realization of our potential.” You could also call it speculative fiction.

About halfway through the film, Holmes heads to prison and the documentary shifts to focus on Billy Evans, the man with whom Holmes has two children. With Holmes in prison, Evans has become her unreliable narrator. He is, as others have noted, also a helpful lens to understand the transformation of Silicon Valley. Evans has seemingly embraced Musk and Andreessen’s brand of vice signaling. We see him shooting guns, extolling Charlie Kirk, and lashing out at people on X under the cloak of anonymity (posting as Holmes, using her X account). He makes grandiose claims about his new company but seems embittered and incapable of submitting to hard questions about it. The veneer of excitement that the couple project in the Theranos 2.0 montage is replaced by a paranoia and resentment that people, including possibly Fielder, might be trying to take him down.

It all made me wonder how Silicon Valley might respond to Theranos’s collapse if it were happening today. Reactionary types in the industry might rally around Holmes, insisting that she was too early or needed more time. They might attack the press more forcefully for trying to bring her down (Holmes and Theranos did try to retaliate against whistleblowers and the reporter who broke the story). Trust in traditional media is even lower now than it was back then, and Silicon Valley has a robust ecosystem of very popular, industry-friendly podcasts and creators who could champion Holmes’s narrative. Evans, for instance, appeared last week on the popular All-In podcast. X is now owned by Musk. The truth takes longer to come out; in some cases, it might feel almost irrelevant. The universal outrage that Theranos sparked in 2015, when news first emerged that something was deeply wrong, sounds quaint now.

“When people ask how she got so far, it’s largely explained by the period she did it in,” Bier wrote in his X post about Holmes. “And she got caught when the world suddenly woke up.” I agree with the first part of Bier’s statement and take issue with the second. The brand of delusion that defined Holmes and the twisted epistemological game that You Can See Everything showcases is still deeply embedded in the industry. None of this is to suggest that everyone in Silicon Valley is engaged in rampant, Theranos-like provable fraud. But today’s entrepreneurs are emboldened to make even more outlandish claims about the future, and it’s more difficult than ever to parse what is real, what is knowingly fake, and what is being offered in good or bad faith.

In response to Bier’s post, Jason Calacanis, an investor and a co-host of All-In, noted that there was “much truth” to Bier’s explanation before he offered a rebuttal befitting a true builder—AI will soon solve the Theranos problem: “In the next ten years, a blood testing product will emerge that will make folks go, ‘See, @ElizabethHolmes was just 20 years too early!’”

The builder is never wrong. If anything, they are just early. The future is coming. Just have faith.

*Illustration sources: Justin Sullivan / Getty; Liesa Johannssen-Koppitz / Bloomberg / Getty; Tayfun Coskun / Anadolu / Getty; USA Today / Reuters.

For the first time since the dawn of the automobile, the majority of new cars are not powered exclusively by gasoline: So far in 2026, electric vehicles and hybrids are outselling traditional gas-powered cars. As oil prices have gone up and up since the start of the war in Iran, drivers from Hanoi to Hamburg are picking up EVs like never before. According to the energy think tank Ember, 49 countries have broken EV-sales records this year. The prime minister of Thailand recently ditched his Rolls-Royce and commuted to work in an all-electric BYD Sealion.

And then there’s the United States. For all the angst over high gas prices, EVs accounted for just 6.4 percent of new-car sales in the U.S. from July to September. That’s a slight rise from January, before the beginning of the war with Iran—but just barely. The U.S. has long been something of an EV laggard, full of drivers (and a president) who are skeptical of battery-powered cars. But even as gas prices have jumped from $3 to $4.50 over the past year, drivers won’t budge.

America’s car fleet was never going to flip overnight. No one forks over $50,000 for a Tesla the moment gas hits $5 a gallon. “If gas prices are higher for a little while,” Elaine Buckberg, a senior fellow at Harvard University’s Salata Institute for Climate and Sustainability and a former executive at General Motors, told me, “people say, ‘Oh, this is going to pass. I’m not going to go and change what vehicle I own.’” But when high gas prices linger for half a year or longer, Buckberg added, the direction of the car industry can shift. The shortages and lines of the 1970s helped kill the mid-century land yacht. In the 2000s, record-high gas prices tempered the craze for gas-guzzling SUVs and pushed millions of Americans to buy hybrids like the Toyota Prius. Now drivers have a real alternative to the whims of the world oil market. They’re just not taking it.

Americans are not all gasoline loyalists who’ve forsworn battery power. In fact, most drivers now say that they are strongly considering an EV as their next vehicle, and the trend is even stronger among people under 35. Pre-owned EVs are surging, but they represent a sliver of the used-car market, because so few of them are out there. The biggest winner from the gasoline spike has been the plain old hybrid, the green-car solution of the George W. Bush years. Sales in the U.S. have jumped 27 percent over the past year.

To some degree, Americans aren’t making the leap for several long-standing reasons: EVs are too expensive, and the options are lacking. Because the U.S. effectively bans electric cars made in China, we are cut off from buying something such as the BYD Dolphin Surf, an electric sedan that sells for about $22,000 in Europe. Many of the models that drivers are familiar with still cannot be had as an EV. Toyotas such as the RAV4 and Camry now come only as hybrids; last year, Ford killed the electric version of the F-150 pickup truck, this country’s most popular vehicle, citing low sales. Ford is now hyping up a roughly $30,000 electric pickup truck, but that won’t arrive until 2027. Meanwhile, America’s only electric minivan is the struggling Volkswagen ID. Buzz, a modern take on the flower-power bus. And although Tesla’s cars have a lot going for them, plenty of Americans say they won’t touch the brand because of Elon Musk. In short, the EV market is full of holes.

[Read: The great Tesla sell-off]

That said, it’s hard not to look at tepid EV sales and wonder whether Americans have lost the plot on electric cars. The price gap between gas and electric cars is closing; more EVs are now available for under $40,000. Charging networks are expanding, and battery ranges are climbing. Compared with even just a year or two ago, buying an electric car comes with fewer trade-offs. And in plenty of ways, EVs are just better. They’re zippy and fun to drive. They don’t require much of the routine maintenance, such as oil changes, that make car ownership annoying. They can do all sorts of things that fossil-fuel burners cannot: refuel at home, keep a house’s lights on during a blackout, store luggage in a “frunk.”

But if gas that’s creeping up on $5 a gallon isn’t enough to persuade people to go electric, what will? When I posed the question to Buckberg, she told me about research she’d worked on while at GM: Drivers were “happy to buy an EV—as long as they don’t have to give up anything they like about their current vehicle,” she said. In other words, it’s not enough for EVs to be good at what they’re good at—especially when relatively few Americans know someone who has one. They need to match the things that gas cars are good at too. Fast chargers are spreading, but not fast enough to free us from the mental load of worrying about where to refuel—something that’s been a given in a nation with a gas station at every corner. And what EVs can’t do yet (unless you’re in China) is recharge in as little time as it takes to pump gas. That’s something American drivers see as a prerequisite, even though most of us don’t regularly take very long road trips and drive an average of 30-some miles a day—a total easily handled by charging at home.

There’s every reason to think that EVs will continue a gradual climb, Corey Cantor, the research director at the Zero Emission Transportation Association, told me. A new wave of hyper-capable electric cars is arriving in America, with ranges above 400 miles per charge. But if EVs really have to do everything gas-powered cars do—and more—before drivers give them a chance, it’ll be a long time before America embraces them en masse. We’re nowhere near the charging speeds that China has achieved nor the density of plugs that would rival the ubiquity of gas stations. Until then, Americans will keep pumping gas into worse cars while the rest of the world has long moved on.

In 1996, during the early years of the web, I could see that those of us who were participating online were creating an extraordinary new record of human activity—and allowing it to disappear almost as quickly as we made it.

The web was becoming the place where government officials made promises, journalists published information in the public interest, and ordinary people shared ideas and documented their lives. Yet it was fragile—equal parts consequential and ephemeral.

Many people assume that once something is published online, it will remain forever retrievable. The saying that “the internet never forgets” has produced a dangerously false sense of permanence. In reality, the internet forgets continuously. Websites are constantly revised, deleted, and abandoned altogether.

This is why, 30 years ago, I founded the Internet Archive. I wanted to build a mechanism for permanence in the internet age. Soon after that, I helped build the Internet Archive’s most-used resource, the Wayback Machine, a system that would preserve webpages as they appeared at any one moment in time, allowing future readers to return to those moments. The idea was simple but urgent: If the internet was going to become humanity’s principal publishing platform, then it needed what every great publishing system that came before it had needed. It needed a library.

People all over the world use the Wayback Machine all the time. What began as an effort to preserve the web has become an essential tool for journalists, researchers, historians, and the broader public.

Yet the future of this tool has become extremely vulnerable. The New York Times is now blocking the Internet Archive’s ability to preserve its articles. The Guardian has restricted our access to its published articles and removed articles from some of the Wayback Machine’s interfaces. USA Today and a growing number of other news organizations have also placed barriers in the path of web archivists (even, yes, The Atlantic). As of May 2026, Harvard’s Nieman Journalism Lab had identified more than 340 local-news sites that were limiting the Internet Archive’s ability to preserve their journalism.

The publishers taking these steps have all pointed to one reason: the unauthorized use of their work by AI companies to train large language models. They complain that our library serves as a back door around paywalls, licensing negotiations, and technical restrictions. The Times has said that it is blocking us because the Wayback Machine allegedly provides “unfettered access” to its work. But this fear-based approach does not take into account that the Internet Archive is taking steps to prevent AI companies from scraping the materials we preserve.

I understand why publishers are angry. Journalism is expensive. Reporting requires skilled people, time, courage, legal support, and institutions willing to withstand pressure. Journalists have watched advertising revenue disappear and subscription growth slow while technology companies build enormously profitable products using material created by others. Publishers have every right to challenge commercial businesses that they believe are inappropriately exploiting their work.

But blocking a nonprofit library is the wrong response to this problem. Impeding the Wayback Machine’s preservation work will not prevent powerful AI companies from gathering training data. It will certainly not restore the advertising-based business model that supported newspapers in the 20th century, a business model that has long since collapsed. What it will do is prevent historians, researchers, courts, ordinary readers—and, most ironic, journalists themselves—from consulting an accurate record of what once appeared on the web. It is difficult to quantify the scale of what we stand to lose. A continuous record of web activity is an enormously powerful accountability tool, especially at a time when seeking the truth and reporting it has become ever more complicated and the informational environment is growing only more complex.

Joe Mullin, who is a senior policy analyst at the Electronic Frontier Foundation, has offered a fitting analogy: Imagine a newspaper announcing that physical libraries will no longer be permitted to preserve copies of its past issues. That is effectively what is happening now. For the past three decades, the Wayback Machine has served as a library for the public web. The destruction of that library is creating holes in the historical record.

A crucial fact in this debate is one that publishers themselves know better than anyone: Journalists rely on the Wayback Machine. They use it to recover deleted tweets, verify past public statements by powerful individuals who wish to avoid scrutiny, examine earlier versions of government documents, and establish what a person or institution said about a subject before some controversy or scandal prompted them to change course. Journalists use the Wayback Machine to see defunct websites and the archives of closed newsrooms. They use it when the primary evidence for a story has vanished from the web.

A news organization that blocks the Wayback Machine may believe it is acting in the best interest of its journalists. But instead it is dismantling one of journalism’s most important investigative instruments.

My argument is simple: The public record needs an independent steward. A society that cannot inspect its past is easier to deceive. In 2017, Katie Hafner and Griffin Palmer, writing for The New York Times, relied on the Wayback Machine to report that private-equity investment was reshaping dermatology practices—expanding the use of physician assistants, increasing the volume of biopsies and procedures, enacting ever more aggressive screening and treatment protocols, and exposing elderly patients to treatments whose benefits were doubtful.

The Wayback Machine gave Hafner and Palmer an empirical record that proved crucial for their story: evidence of change over time. The current website of Advanced Dermatology and Cosmetic Surgery (one of the companies they were looking into) showed a sprawling operation with more than 180 locations and 124 physician assistants. But earlier versions of the site showed that the number of physician assistants listed by the company had increased by 400 percent since 2008, amid a debate over whether physician assistants are qualified to properly make diagnoses. When I spoke with Hafner about the story, she told me she could not have made her point as powerfully as she did without the Wayback Machine. The records she retrieved from our archive enabled her to demonstrate the scale, speed, and direction of a corporate transformation, including the rapid growth of a business model that placed many more lesser-trained practitioners in clinics serving vulnerable patients.

Hafner and Palmer are among thousands of reporters whose use of the Wayback Machine has contributed to extraordinary journalism. A politics reporter may find that an agency quietly removed a document, or even a single sentence, in the face of criticism. A city-hall reporter may need an earlier version of a municipal page containing budget figures, meeting records, and public-health guidance. A technology reporter may need to compare successive versions of a platform’s terms of service. An investigative journalist examining allegations against a public figure may need to scour a decade-old webpage that has since been deleted.

Government agencies routinely scrub important information from their sites, and the Wayback Machine is a faster and more practical method of discovery than a Freedom of Information Act request. These are not hypothetical benefits. Over and over and over again, the Wayback Machine has confirmed facts that would be irretrievable without an independent archive.

Our own monitoring finds that more than 100 news articles in a typical month cite or rely on material preserved by the Wayback Machine. We can reasonably assume that there are tens of thousands of additional visits to the Wayback Machine every month, when an archived page might help illuminate the truth.

Mark Graham, who directs the Wayback Machine, recently described having visited The New York Times to meet with data journalists and reporters—and being approached by a senior researcher who thanked him profusely. The researcher said Times staffers used the Wayback Machine constantly and had even found material there that they could not find anywhere else. Consider what that means. Even one of the best-resourced news organizations in the world does not possess a complete, easily usable record of the material it needs.

Plenty of journalists—including at The Atlantic—have championed and defended the importance of the Wayback Machine. MS NOW’s Rachel Maddow has called the Internet Archive a national treasure and said she uses it daily. She and hundreds of other journalists have signed a statement calling for web preservation.

These journalists understand that their work is cumulative. A story links to documents, earlier articles, public statements, data sets, and numerous other verifiable facts in the public record. Future stories return to (and build upon) that evidence. When underlying links die, one of the chains of verification breaks along with it.

In 2024, the Pew Research Center announced the findings of a study that tracked 1 million webpages over a 10-year period. At the end of the study, one-quarter of those pages were no longer accessible. Among pages that existed in 2013, the first year of the study, the loss rate was 38 percent. Even among pages collected as recently as 2021, roughly one in five had disappeared within two years.

This is not the timescale of ancient history. But it is often the timescale of someone’s political career, a long-running lawsuit, or a company’s rise and fall.

Elements of the web disappear for all kinds of reasons, nefarious and mundane. Perhaps a page is deleted during a redesign. Maybe a publication changes ownership, or a content-management system is replaced.

After the Pew study came out, the Internet Archive took a closer look at its data set to determine how much of the lost material had been preserved by the Wayback Machine. The analysis found that 16 percent of the vanished URLs are accessible via the Wayback Machine.

Consider the Rocky Mountain News. Colorado’s oldest newspaper published for nearly 150 years and earned a reputation for forceful accountability journalism that exposed corruption, misspent public money, and institutions that were failing the people they were meant to serve. In its final decade, the paper won four—I repeat, four —Pulitzer Prizes: for photographic coverage of the Columbine school shooting, for its images of Colorado’s devastating wildfires, and for the writing and photography of “Final Salute,” a deeply reported account of Marines killed in Iraq. In 2015, The Atlantic’s Adrienne LaFrance reported the story of how one of the paper’s Pulitzer finalists successfully resuscitated an investigative series of his that had vanished from the web when the paper did. Her story concluded with a link to the series that he’d saved. Even that link is now dead. (Though you can still find the series here.)

No institution, however prestigious, is immortal. The Wayback Machine holds the work of publications that no longer exist, such as the Rocky Mountain News. It preserves the archive of MTV News and BET.com, including decades of cultural reporting removed from the web without warning. It preserves a robust record of early-2000s web culture from Gawker Media. It provides access to work produced by Apple Daily and Stand News, Hong Kong–based news outlets that were forced to shut down in 2021 as part of a government crackdown, as well as independent Russian-language journalism, some of it produced in exile. It holds the only copy of USA Today from September 11, 2001, because the paper’s website was overwhelmed and its online editors continuously updated stories by hand without preserving the changes made along the way.

The Internet Archive is not an AI company. A commercial AI laboratory, a search engine, a research project, and a nonprofit library may all use automated systems to retrieve publicly available webpages. From a server’s perspective, each may appear to be a bot. But not all bots are bad bots. They do not have the same mission, business model, or conduct.

Publishers’ concerns about abusive scraping are merited. I know this acutely because the Internet Archive has itself been subjected to extreme machine traffic. In one incident, an AI startup sent tens of thousands of requests per second and overloaded our systems. We blocked the offending hosts, contacted the company, and worked to stop the behavior.

When a company attempts to extract material at scale, it burdens our nonprofit infrastructure and interferes with access by ordinary users. We understand why publishers object to how AI companies have used their work. But recognizing a problem does not justify attacking the wrong institution and disregarding the public interest. In short, the Wayback Machine has become collateral damage in a fight between publishers and AI companies. Blocking the Internet Archive will not stop AI. Commercial AI companies have crawled the live web, licensed large data sets, purchased access from platforms, obtained third-party collections, and built their own retrieval systems. Publishers can sue those companies, negotiate with them, or build technical controls specifically directed at them. We have already built such controls to protect the material on our site—including the material we preserve from news organizations.

Web archiving belongs to a well-established legal tradition. Courts have repeatedly recognized that copying material for indexing and search can be a transformative fair use. A searchable system cannot function without making technical copies of the material it indexes. In the now-famous Google Books litigation, in which authors sued Google over the digitization of millions of books, courts held that scanning books to create a searchable research tool served a different purpose from selling or reading the original books, and was fair use under U.S. copyright law.

Defending preservation does not require ignoring publishers’ legitimate concerns. The Internet Archive has adapted repeatedly as the web has changed. We use rate limits, filtering, monitoring, and network-security systems to respond to abusive access. We have restricted bulk functionality in particular contexts and optimized important parts of the Wayback Machine for human use.

Those efforts can and should continue. Publishers and archives can distinguish ordinary human consultation from sustained bulk extraction. We can develop better methods for identifying abusive traffic, create responsible-access arrangements, improve communication during high-volume incidents, and discuss access to application programming interfaces separately from the preservation of publicly available pages. Most important, we can preserve records while applying reasonable protections against mass downloading.

The press cannot have it both ways. Journalists describe their work as the first draft of history—but a draft of history must survive if it is to be revised, improved upon, and understood.

For the past 24 hours, Silicon Valley has been obsessed with a 26,000-word blog post by the AI researcher Nick Cammarata. The document has been called “the most important essay,” “the most important result,” and “the single best piece of news” of the year, or in the history of AI research—or, perhaps, ever. Even OpenAI CEO Sam Altman lauded it. Cammarata makes a series of bold, almost astonishing claims about the nature of AI that, if true, would upend some of his industry’s basic assumptions about the very thing it is building. And after months of AI-hacking sprees and mounting public backlash against the industry, many people see Cammarata as offering a glimmer of hope when it comes to not just understanding but also controlling the technology being unleashed on the world.

Boiled down, Cammarata’s main contention is that generative-AI models are not actually black boxes, contrary to the common perception that these systems are so big and labyrinthine as to be nearly impenetrable. In fact, Cammarata argues, there is a relatively clear path toward understanding them. If that’s the case, the material significance could be tremendous: It would provide a way to map AI models’ inner workings and, in turn, perhaps prevent them from wreaking all sorts of harm—autonomously hacking companies and sovereign nations, encouraging users’ delusions, appearing to deceive AI researchers themselves. There is an entire field of research, called “mechanistic interpretability,” devoted to the study of just this.

The argument is just as important as the person making it. Cammarata, a former OpenAI employee, is a respected figure in his field. Years ago, he played a key role in establishing the opposite claim, which is now common wisdom: that neural networks, the software architecture that generative AI is built on, are impossible, or at least very difficult, to understand. In this new essay, Cammarata, who now works at the AI-research company Goodfire, essentially says that he and his entire field have been wrong ever since. “The claims here are viscerally shocking,” one observer wrote. Or, as another wrote on X, they are “big big big big.”

[Read: Imagine every AI agent on Earth buying the same cheap paper towels]

Neural networks, inspired by the human brain, consist of layered connections among individual nodes, or “neurons,” that activate and send information to one another when responding to a query. Earlier research, including Cammarata’s, established that individual machine neurons can activate in response to many different, unrelated topics—for instance, one neuron can correspond to images of cat faces, cat legs, and the front of cars. Nobody was sure why this happened, but one explanation was that, because AI models contain far more information than they have neurons, they have to repurpose these nodes. Regardless, the basic units of an AI model having seemingly arbitrary, multitudinous functions would make it very difficult to understand what is going on under the hood.

Cammarata now thinks that individual neurons can be understood and that each has a self-consistent function. Neural networks are not “tangled knots,” he wrote, but “clean.” He found that the cat-face and car-face neuron, for instance, was actually responding to things with a V-shape—including V-shaped flames and jets of water. This would make AI’s internal mechanisms relatively straightforward and predictable rather than mysterious and nebulous. And if we can harness that understanding, some people in the industry believe, AI will be much less likely to go haywire and, say, kill us all. “This is truly one of the greatest cases for optimism in years!” the OpenAI researcher Adam Majmudar wrote on X.

Not everyone agrees with Cammarata. For one, his 26,000 words were merely a “working hypothesis” for a “forthcoming technical paper” and were thus hard, on a technical level, to evaluate. And the neural networks he studied most closely were vision models, which are generally simpler and better understood than the language models underlying ChatGPT and Claude. Cammarata could be mistaken, or he could be overstating his case. The blog post isn’t peer-reviewed, and the forthcoming paper probably won’t be, either, as is typical in the field. The research itself is something of a black box. (Cammarata, who did not immediately respond to a request for comment, wrote on X earlier today that he appreciates all of the discussion and that his approach “seems promising so far” for “bigger modern language models.”)

Discussions of Cammarata’s essay have revealed philosophical rifts. “This is interesting work, but I don’t think it’s a significant update to the big picture,” Neel Nanda, who leads mechanistic interpretability at Google DeepMind, wrote in his appraisal of Cammarata’s findings. Even if researchers could approximate the function of every neuron in Claude Mythos 5.1 or GPT-6 Astra, that’s not the same as understanding how they connect and how the network works as a whole. Physicists have good models of atoms and interatomic interactions, but they cannot from first principles explain or predict every physical event (hence unpredictable extreme weather or, for that matter, the notion of free will). You know all of your friends from school, various workplaces, and the local softball league, but you can’t always know whether they’ll get along at your birthday party. More to the point, insights into the inner workings of an AI model will not necessarily produce insights into how to make the technology more reliable. Interpretability may be necessary to build safer AI models, but it is by no means sufficient.

If generative-AI models are actually legible under the hood, that would undoubtedly help make the technology safer. But many people in Silicon Valley, and especially employees at OpenAI, are discussing Cammarata’s blog post with a breathless enthusiasm that echoes Silicon Valley’s long track record of easy technical solutions and grand promises—an attitude that belies its mixed record on following through.

Any research finding that pertains to AI safety will require an industry chasing multitrillion-dollar public offerings to slow down and take a long, hard look at itself. Perhaps even more concerning than an unintelligible superintelligence is the fact that a group of companies openly state that they don’t really know what they are building—and then continue to push that technology to extremes anyway. If Cammarata is correct, AI companies will have to work with one another, outside auditors, and governments to figure out how to translate such a breakthrough into safe and reliable products. And right now, absent any serious regulation, nobody can decide to take such action other than the companies themselves. All of which suggests that the real black boxes are OpenAI and Anthropic themselves.

Close your eyes, and imagine a frictionless day.

You wake up to a notification that a grocery-store employee will deliver all of the ingredients for the three recipes you saved last night to your “Weeknight Dinner” folder. While you brew your coffee, a program browses online retailers to compare toilet-paper prices, and another solicits competitive bids for Sheetrock work you need done in the garage. During your commute, a bot moves your upcoming dentist appointment to avoid a scheduling conflict. On your lunch break, you get an alert that computers acting on your behalf noticed some unclaimed funds and class-action payouts in a state database and claimed them. Also, the machine has negotiated a cheaper internet plan. By the time you get home, the groceries are on your doorstep, and you’ve gotten back $150.

This is the future that OpenAI, Meta, xAI, Google, and the rest are selling with a spate of always-on AI assistants that work for you. “Imagine every person on earth with a second mind beyond their own,” Meta’s chief AI officer, Alexandr Wang, wrote of the company’s new AI helper, Muse. A few weeks after Muse’s launch, OpenAI launched its competing product, Dots, which Sam Altman claimed would let you “get more of your time and attention back.”

[Read: The singularity is not what it seems]

These AI agents promise to bulldoze the inefficiencies of daily life. Think about all of the opportunities you’re currently leaving on the table: You don’t change cell-service providers for a flash sale because of the paperwork involved. You haven’t fiddled with your 401(k) investments in years. You have kids to pick up, a date to dress for, sleep to catch up on. But your AI agents are different—they’re relentless, tireless, and shameless. They work in the background, without supervision, eagerly helping with all of these tasks and a bunch you didn’t even ask for. They use every tactic and loophole at their disposal to get you the best price and the highest-quality service because they’re able to pay attention to everything at once.

The AI companies see a massive opportunity for hyper-personalized agents to become a new intermediary layer for the internet and the physical world. This future is already unfolding: People are giving Muse access to their phone number, email, credit cards, and bank accounts to make dinner reservations, contact customer service, and book vacations. Mark Zuckerberg says that he is using Muse to coach his martial-arts form and help him bake with his daughter (he isn’t very familiar with baking, being “normally into smoking meats”).

Silicon Valley’s vision of boundless efficiency is shortsighted, or at least incomplete. Near the end of his blog post about Muse, Wang writes that “we’ve never seen humanity with every single person’s agency fully switched on.” He paints a utopian vision of laptop-class optimization. But such frictionless commerce, if it ever does come to pass, could bring about something else entirely: chaos.

The era of the agentic web started less than 24 months ago. Anthropic’s Claude Code, introduced in early 2025, allowed agents to write software and use other programs on your laptop. OpenClaw, launched later that same year, could navigate the web on your behalf. In the spring of 2026, Instinct, an invite-only app that let users text requests to a bot over iMessage or WhatsApp, began capturing the imagination of Silicon Valley insiders. It could book museum tickets, purchase gifts, and find in-network doctors.  

But it wasn’t until the launch of Meta’s Muse in early September that personalized agents hit the mainstream. Muse has been downloaded more than 5 million times in the past month and is especially popular among early adopters—entrepreneurs, investors, software developers. These users are now reporting back on their experiences and offering previews of our potential agentic future, which sounds both strange and jagged.

When we reached out to Rajan Lukose, a start-up founder in the Bay Area, to discuss his use of AI assistants, he had his Muse set up a calendar invite. In the late 1990s and 2000s, Lukose worked for corporate research-and-development teams in Silicon Valley and did early research on web agents. Now, he told us, “they’re actually, finally becoming sort of real.” Lukose has been using Meta’s program for “annoying things that I would never do”: tracking travel sites over time to book the cheapest flights, submitting receipts for DoorDash credits, sorting out problems with a magazine subscription. “That’s something that’s a real pain in the ass that just gets done,” he said.

Lukose said that the primary upside of using an agent is clawing back time and money. A few years ago, for instance, Lukose installed an automated tool that went through his Gmail, looked at various transactions, and filed claims under price-purchase guarantees. “That was like $600 that I just would not have gotten myself,” he said. The agentic era promises to deliver those kinds of savings to everyone.

Decades ago, the internet was supposed to create perfect markets, with users able to compare the cost and quality of products and services from anywhere in the world. But it failed: People continued to make lots of decisions based on habit. That, in turn, is good for business. “A lot of things online are premised on limited human attention,” Andrey Fradkin, an economist at Boston University, told us. Digital advertisements, recommendation algorithms, brand reputation, and celebrity endorsements steered consumers toward making suboptimal purchases.

AI agents are now here to finally fulfill the promise of perfect commerce. “The agent knows your preferences; they can go to every single platform, find the best products, the best deal for you, and then help you make that purchase,” Fradkin said. The agent won’t be fooled by sponsored posts or give up on shopping when they get distracted or hungry. “The current attention-based economy—if you gain the eyeball, you have a better chance to sell certain goods—that may change,” Ken Huang, who studies agentic AI at the University of San Francisco and runs an AI consultancy, told us. The bots are impervious to human-targeted nudges. Their attention is infinite.

In the short term, however, this automated quest for efficiency is creating some unsettling dynamics. Over the past month, X, Reddit, and other forums have been overrun with examples of agents haggling customer-service representatives and contractors. Take the case of the Muse user Luke Whiting, who sent his agent to check on a prescription at Walgreens. Whiting, who works for an open-source-database company in San Francisco, told us that the pharmacist at the other end of the call refused to speak to an AI system. His Muse suggested that the best way forward would be for Whiting to abandon using the agent altogether: “Honest take: other stores might say the same thing,” it said. Having the Muse call other pharmacies “could just burn time. The fastest path is probably you calling yourself.”

On a Subreddit for call-center employees, a recent thread is titled, in all caps, “I AM NOT AI.” “I have been asked 3 times this week ‘are you human?’” one person replied. Other users who identify themselves as call-center workers discuss their frustrations. One says that when customers question whether they’re talking to AI, they like to respond, “I’m a real boy,” like Pinocchio. Another, fed up with rude callers, vented, “Customers deserve ai.”  

Companies, meanwhile, are starting to tap AI systems to squeeze more out of their customers. According to Fradkin, many hospitals don’t charge as much as they could, simply because medical-billing codes are countless and labyrinthine. But using AI to document and code patient visits reliably allows hospitals to charge for more, and more expensive, services. That likely contributed to patient costs increasing by $650 million across Blue Cross Blue Shield from 2023 to 2025. And that’s only the beginning; health payments are projected to rise another 9 percent next year, largely because of AI billing. McDonald’s is reportedly using AI to analyze millions of consumer transactions and adjust prices at each location, for each menu item, based on local demand. (McDonald’s has said that its pricing portal is “a tool, not a mandate” to help franchisees.) It’s not hard to imagine vendors using AI to develop personalized, on-demand inventory management and dynamic pricing to maximize profits for streaming subscriptions, sports tickets, drug prescriptions, consumer electronics, insurance policies, and loans.

This optimization assumes that the agents work as intended. So far, there have been plenty of stumbles: In a paper published earlier this year, a group of AI researchers found that agents, given assistant-style tasks in a virtual environment, leaked Social Security numbers and were easily hijacked by other people. Agents powered by Chinese models refused to complete tasks that involved politically sensitive topics. Two weeks ago, a tech executive’s AI agent randomly posted his financial information in a workplace chatroom for employees to see.

Things could get weirder still: Tech companies market their agents as perfectly rational, when, in actuality, they can be extremely erratic and unpredictable. Sophia Kazinnik, who studies AI agents at Stanford, found some models to be overly generous and others to be exceedingly greedy. Across the board, they were far less patient than humans. When two personal AI agents are haggling over prices in the wild, the outcomes might not be maximally profitable. And there’s no assurance that all future agents will be created equal: More powerful models may have an upper hand in negotiations.

Even if the agents are able to effectively shop and negotiate for us, important rules governing the privacy rights, liability, and economic allegiance associated with these tools will have to be codified. Does your Muse have a fiduciary responsibility to you, or a legal responsibility to Meta? If Muse pays for a $2,000 espresso machine instead of a $200 one, can you be refunded? If Muse leaks your account and routing numbers, is Chase required to refund any fraudulent payments?

At some point, everyone’s agents could reach parity. “Just like today, every business has a website and social-media accounts and email addresses and all these things, every business is going to have agents too,” Zuckerberg said in an interview this week.

In this future, communication and commerce function almost entirely via computer proxies. Your agent calls a fancy restaurant for a dinner reservation, and a digital maître d’ picks up. Your agent and your insurer’s agent and your hospital’s agent all negotiate the price for your recent appointment. Prices fall, buyers and sellers are better matched, and businesses are incentivized to deliver higher-quality, customized services and products. This is “what economists dream about,” Fradkin said.

But friction can serve a purpose for businesses. Banks make great sums when customers leave money in low-interest checking or savings accounts, and studies show that sizable portions of retailer subscription revenue are driven purely by “inertia,” another way of saying that people forget to cancel. (Friction can help consumers too: Inattention and incomplete information at times lead landlords, pharmacies, and other businesses to undercharge.) Our entire economy and basically every company are structured around the notion that humans have mental and physical limitations. Without them, existing market structures could collapse.

“There’s this strange protection provided by the inefficiencies in our economy,” Rory Van Loo, a lawyer and a professor at the Wharton School, told us recently. In 2019, well before the rise of large language models, Van Loo envisioned a phenomenon he dubbed “digital market perfection.” Imagine everyone has a digital agent that is always looking for a lower price or a newer, better service. In this world, Van Loo argues, agents could, all at once, direct millions of consumers toward a new product or service—a process he calls “hyperswitching.” That would in theory be good news for consumers, but it would be sheer hell for companies. The nightmare scenario that Van Loo and others have conjured is a style of agentic bank run, where similarly attentive agents all pull money from a bank at once to get their clients better interest rates elsewhere.

Maybe this sounds good or exciting to you—a kind of consumer’s revenge—but in practice, stable commercial markets would begin to look and function like the stock market. “If, all of a sudden, people could purchase whatever they wanted at the lowest price with the perfect ability to see every possible option, there’d be a lot of volatility in the world. It would seem highly efficient from an economic standpoint, but it could have devastating effects,” Van Loo said. Picture 1 million personal agents all switching their consumers to a new brand of paper towels on Amazon because a markdown made them three cents cheaper. The same Sweetgreen salad is $18 one week and $38 the next. Your electric bill is $150 one month and $450 the month after, and you’re uncertain with each payment whether your local utility is about to go bankrupt.

“Stocks are bought and sold fast; markets change and plummet—such that we have a stop button for it, because it can go off the rails,” Van Loo noted. “But we don’t have those kinds of protections in most other markets.” There’s potential for collateral damage everywhere. If millions of people switching to higher-yield accounts causes a bank run, that’s a disaster for any person or institutional investor with money locked up at that bank. If half of Verizon’s customers switch to AT&T in one week to take advantage of a flash sale, the business could collapse overnight, affecting the other half of its customers. Sometimes, this rapid switching could signal healthy competition. Other times, it could cause organizations to implode.

A world where AI proxies do the planning and talking could also mean a future where humans are constantly surprised by things they once were able to anticipate: a birthday present for your spouse ordered in a preposterous size that you must try to explain; a plumber showing up out of the blue.

Or perhaps none of this will happen. It’s possible that an assistant-dominated economy turns out to be far more mundane, defined by tacked-on fees to keep the agents in check. Instead of agents allowing you to negotiate and bring down health-care, insurance, and electrical bills, the companies responsible for all of these services charge your agents every time they request to pull your data. Banks, to bolster their revenues in the event of a bank run, institute new and higher fees on checking accounts. Amazon, unable to profit from dynamic pricing and ads in the face of perfectly rational agents, raises prices for consumers and takes a bigger cut from manufacturers. “Generally, if one fee goes away, you get another,” Fradkin said. In other words, everything could be faster, easier, and more efficient for the consumer—but potentially more expensive.

Van Loo said that the bigger AI companies will eventually find ways to strike deals with platforms via these protocols, negotiating access, commissions, and partnerships in ways that benefit the companies, not necessarily the consumers. This could result in a sort of preferential treatment like the deals already in place between Meta and Walmart, Sephora, and other retailers, which will effectively cause Muse to prioritize their products. Taking another approach, Amazon has blocked Muse and encourages visitors to browse with its own agent, “Alexa for Shopping,” meaning the company is basically negotiating on both sides of the deal. Meanwhile, this week, JPMorgan announced that it would not allow Instinct, Muse, or other agents into customer accounts. Ultimately, this might all lead to “the same old heavily obstructed, inefficient world we have now,” Van Loo said. “But it will look far more efficient, because it will save some people time and feel much more convenient.” Very powerful companies will design new structures for markets and the internet, determining what can be sold, by whom, and for how much—as they always have.

For all the time and money the agentic future could save us, the technology entails another sort of cost: an additional layer of abstraction and disengagement. We’ve largely outsourced our decision making to a small number of platforms: Amazon guides our shopping; Google Maps steers us toward certain restaurants and primary-care providers; Instagram dictates our news consumption. We are told that these recommendations and ads are, based on troves of personal data, already optimal. The companies pushing agents assure us that the decision making will now be even more rational and the deals even better. But checking these claims for ourselves will be harder, because the logic behind those recommendations will be buried in large-language-model chains of reasoning. Eventually, the skeptical instinct to check agents’ work could atrophy altogether. Over time, humans will play a smaller and smaller role in their own life.

In the age of the algorithm, it’s easy to throw out cliché terms such as late-stage capitalism in an attempt to describe the extractive, inhumane dynamics of the economy. But what is an agentic future if not the furthest extension of our current market logic? A future of capitalism so perfect, so efficient, that it doesn’t need us at all.

For today’s high schoolers, a typical school day from a decade ago must seem like something from the Dark Ages. These days, students are issued digital textbooks instead of physical ones, they learn how to conduct chemistry experiments through apps that simulate a science lab, and they take standardized tests on a Chromebook. All of this is the result of many years of frenzied investment in ed tech, capped by a pandemic that forced virtual learning. The overwhelming majority of schools nationwide provide each student with their own laptop or tablet.

All of these screens have changed the dynamic of the classroom. When a student is supposed to be using their school-issued iPad to diligently take notes during math class, they might instead be messaging a friend. Across the country, some parents and teachers are mounting a backlash against screens in the classroom. In April, the Los Angeles school board voted to stop giving devices to its youngest students and implemented screen-time limits for all grades.

But other schools are pulling back on ed tech for an entirely different reason. Chromebooks, like essentially all other consumer electronics, need RAM, or random-access memory. So do the data centers that power every conversation with a chatbot. The AI industry is using so many memory chips that it’s led to a global shortage, making gadgets much more expensive. For an individual consumer, the memory crunch—RAMageddon, if you will—means having to spend a few hundred dollars more on a new computer. But school districts, many of which are already cash-strapped, buy hundreds or thousands of computers every year. The administrators I spoke with for this story told me that they have had to scramble to acquire all the electronics they now need. As RAMageddon drags on, they said, something has got to give.

[Read: If you need a laptop, buy it now]

Schools started picking up on the memory shortage last winter, months before most of the country. They have to figure out bulk orders for laptops and tablets long before the academic year begins. Many of them turn to third-party vendors, which have typically provided quotes that are locked in for a month, giving technology directors time to secure the funds from their districts to make purchases that can run millions of dollars. As memory prices have soared, Mark Samberg, the chief technology officer of Chatham County Schools in North Carolina, is hearing a different message from his vendors. “Now quotes are going to be good for 48 hours, a week, or 14 days,” he told me.

Samberg scooped up the Chromebooks he could find. Other districts have been less lucky. In March, a Chromebook vendor warned Homewood-Flossmoor High School, in Illinois, that its prices would be going up soon. Steve Richardson, the district’s IT director, quickly placed an order, only to be eventually told it couldn’t be fulfilled, because memory chips were on back order. “We were kind of in dire straits,” Richardson told me. In 2023, Chromebooks cost the school $353 apiece. The price rose to $400, then $412, and then $427. The most recent batch that Richardson purchased, in August, cost $508 each. But the district needed them to ensure that every student had a Chromebook ahead of the PSAT, which is now entirely digital.

So far, schools have been willing to stomach the costs to continue giving every student a device. Districts initially decided to implement “one to one” policies on the premise that technology can be an equalizing force in the classroom; many educators and administrators still believe that. Dorann Avey, the board chair of the State Educational Technology Directors Association, told me that she is concerned that RAMageddon will widen the digital divide. Devices are not for “just writing papers,” she said. “We’re trying to prepare students to learn different software that is workplace- and workforce-appropriate.”

That has meant making difficult trade-offs. Because of the extra spending on laptops for students, Homewood-Flossmoor didn’t buy any new smart whiteboards or computers for teachers this year. In Texas, the Paradise Independent School District has considered asking students to bring in their own laptop or tablet, Director of Technology Doug Bryant told me.

A common solution has been to extend the lifetime of a laptop. Instead of replacing a device every four years, schools are opting for seven or eight to reduce how many units they need to purchase in a given year. But that introduces new problems: Older computers sometimes aren’t able to update to the latest operating system, which could create cybersecurity risks. Schools are already a top target for hackers, because they store lots of sensitive personal information but have limited resources. Plus, as a device gets older, its battery degrades and may ultimately need to be replaced. That costs “time and money,” Avey said. “But it’s still less money than buying brand-new devices.”

All of this may not be enough. Multiple administrators told me that they are already considering rolling back their one-to-one device policies because of RAMageddon, and the challenges for schools are poised to mount. Heading into this school year, some districts were able to skirt the worst of the memory shortage by buying Chromebooks before prices really started to soar. Aaron Smith, the CTO of Loudoun County Public Schools, in Virginia, told me that he sought special permission from his school board to commit to purchasing nearly 24,400 Chromebooks in February instead of May. Memory prices jumped so quickly over that time span that the decision ultimately saved the district $1.66 million, he said. But Smith and other technology directors may not have the same luck as they begin looking to purchase gadgets for the next school year. An executive at Micron, one of the world’s largest RAM manufacturers, recently warned that there’s “no line of sight” to the end of the memory shortage. If schools wait to buy, prices could end up going up even more. But buying early might mean devices will languish in storage for months as the warranty clock ticks away.

RAMageddon will force schools to scrutinize which tech purchases actually improve learning, and which don’t. The fallout will reach far beyond Chromebooks. “We’re not going to be able to afford the servers that we thought we were,” said Samberg, the Chatham County Schools CTO. Those servers power functions such as security cameras, badge readers, and account logins—and every one of them needs RAM.

A man walks into a restaurant. He doesn’t have a reservation, but that shouldn’t be a problem; there are ample seats at the bar, visible just over the hostess’s shoulder. “Of course,” she says, smiling. “We can seat you right away.” As she gathers menus, he absorbs the pleasant hum of a busy dining room, breathes in the tantalizing smells from the semi-open kitchen. Then, the hostess stops, Columbo-like, to interrupt with one more question: “And what’s your phone number?”

The man was me. The restaurant was the Kingsway, a newish, Michelin-recommended spot in New Orleans. I’m not going to give you my phone number, but I gave it to the hostess, almost without thinking. And then, I did think about it: a lot. Throughout the meal, and for a long time afterward. There had been no reservation to check, no rewards program, no waiting list. So why the automatic ask?

Let me say quickly that I liked the Kingsway very much. It’s coolly fancy, with a pan-Asian menu evoking the journey of the chef-owner Ashwin Vilkhu’s family from South Asia to the Gulf South. The result brings to mind an earlier era’s scruffy raiding of the global pantry, but this time executed with thought, technique, and precision. As I tucked into a tangle of Singapore-style chili-crab noodles, the term stoner soigné passed through my head.

But that very first interaction with the hostess colored my meal. I brooded. What if I had just wanted to sit at the bar anonymously? Was that not one of life’s great pleasures? (I was working up a head of steam.) Was there no place in this enshittified world where a man could get a break?

[Read: You will never get that restaurant reservation]

It mattered little that I knew the answer. Like pretty much everyone else, I am well aware that my information is strip-mined wherever I go, in more ways than I can count and in many that I will never know. Like everyone else, I spend my days downloading apps to operate things such as toothbrushes, unsubscribing from marketing lists by somehow giving up more personal information, rejecting cookies like I’m on a gluten-free diet.

Nor am I overly romantic about restaurants, which I know have been collecting data for years. The reservation platform OpenTable began making algorithmic recommendations to users as early as 2014; by 2017, The New York Times reported that Big Data was “staging an intervention” to save the restaurant industry, which was suffering from stagnating growth and flattening profits.

Much of the data collection has been justified in the name of customer service—a way to track allergies, preferences, birthdays, and so on, thus bringing technological efficiency to a job previously handled by a heavily notated reservation book or the encyclopedic mind of a gifted maître d’. “We try to build a database on everybody and get as much information as possible,” Vilkhu told me when I called him a few weeks after my dinner. His restaurant uses OpenTable, and he knew, for instance, which bartenders had served me, the friend who had joined me, that I had complimented the cocktail list and enjoyed the crawfish dumplings. This information, he assured me, was shared with the Kingsway’s sister restaurant but nobody else: “We just want to be able to curate the experience as best as we possibly can.”

I believed him. But I was nevertheless a little shocked to go to the websites of the big reservation companies and see how they sold the impulse to care for guests to restaurants (which, in case it’s not clear, are their customers). At best, the sales pitches were dispiritingly cold, with talk of CRMs, GRMs, CDPs, and integrated tech stacks. At worst, they spoke of “capturing” guests’ data, “fostering loyalty,” and “upselling 24/7.” The “AI powered hospitality platform” Loyalist promised “a 360° view of every guest across dine-in and event spend,” itself a sentence seemingly powered by AI. The sales decks were like those optical illusions that turn from adorable to sinister when hit with different colored lights, flipping the language of care and hospitality to the language of surveillance and theft.

Most striking was OpenTable’s promotion of a feature, currently in beta, called Global Insights. Until now, as Vilkhu had told me, whatever data a restaurant collected through the reservation platforms remained confined to that restaurant or restaurant group. Global Insights, which is available to OpenTable Pro subscribers, is the first feature to jump that garden wall and allow other restaurants access to the same information—in OpenTable’s words, to “get an idea of how your diners behave at other restaurants across the OpenTable Diner Network.” With an account, any restaurant can identify big spenders, diners who tend to linger at a table, diners who order a lot of drinks, frequent late-cancelers, and chronic “no-showers,” which I initially read as being about personal hygiene.

[Read: Please don’t make me download another app]

My outrage at this new invasion of privacy was accompanied by a pang of embarrassment: This is one of those times in history when it’s easy to get old fast, to wake up one morning and find yourself appalled by norms that everybody around you seems to take for granted. (For instance, the notion that one’s phone number remains a piece of intimate information.) Still, the idea of a restaurant I’d never been to using a dossier on me to decide, say, where I should sit, what I was likely to order, even whether I should be given a reservation at all, gave me chills.  

Curious to see what insights on me OpenTable might share with that hypothetical restaurant, I sent the company a request for my personal data. To its credit, the company makes this easy with an online form; less than 24 hours later, a zip file arrived. (Resy requires a request by email; a week later, I was told that it was “under review” and have heard nothing since. SevenRooms, another reservation platform, says that it is merely a processor of data and suggested that I contact individual restaurants.)

The OpenTable file was 48 pages, in the monospaced typeface of computer code. Much of it was indecipherable, aside from dozens of instances of my phone number (aha!) and semi-redacted credit-card information. Despite its length, the document contained only a fraction of the reservations I’ve made over the years. The Kingsway wasn’t in there; neither was the first OpenTable reservation I can find in my records, at Zenkichi, a Japanese restaurant in Williamsburg that I visited in 2008. The file had a note that read “GIN,” which I grant was on the money about my inclinations, but it wrongly identified my home market as Austin. If this was truly the sum of OpenTable’s knowledge, I had less to immediately worry about than I’d anticipated.       

[Read: Restaurants are the new factories]

At the same time, seeing the traces of past meals emerge from the sea of characters was eerie. Many I remembered; plenty I had not thought about since. Next to many of the restaurant names was appended a note: “Permanently Closed.” Lowlife, Church & State, Hunky Dory, La Mangeoire—all no more. Others didn’t have the tag, but I knew they were gone too. Like ghosts, they swam through the mists of the OpenTable servers, rattling chains forged of bad leases, unruly food costs, fickle trends, misbegotten dreams, and a thousand other maladies. I could almost hear them wail, “Doooooon’t judge ussss … We just wanted to stay aliiiiiive …”

Indeed, restaurants are often asked to bear a moral and behavioral burden not generally placed on other small, low-margin, independent businesses. The more the public has learned about how restaurants work, the more they have scrutinized the ways in which they handle all sorts of social issues, such as racial and gender equity, climate change, and, yes, data privacy. These are all great and worthy considerations, but they do illustrate a gulf between how some customers (and many food writers) see restaurants and how many chefs and restaurateurs do: The former might like them to, you know, just solve the problem of ethical consumption under late-stage capitalism, while the latter are preoccupied with how to lower the cost of a lamb chop by 20 cents so they can pay that month’s electric bill. In the bigger picture, both diners and restaurateurs are reckoning with the same systemic forces.

But the question remains. Knowing all of this—not to mention with a plate of crawfish dumplings before me—why was I so rankled that night when asked for my number? Maybe because it’s exhausting, these constant micro-negotiations between dignity and grift. Exhausting, and lonely: Most people know that completely opting out of data collection isn’t an option, short of walking into the woods. So we’re left to improvise which of the numberless small violations are worth summoning the energy to resist: when we’re willing to have an awkward interaction, how embarrassed our companion will be, which employee might get in trouble, whether making a scene is worth the hassle. Like so many things, we’re on our own.

And maybe I am romantic about restaurants. About a 10-minute drive from the Kingsway is another restaurant, Crescent City Steakhouse, which opened in 1934. It features four booths along one wall that are outfitted with curtains that can be drawn for maximum privacy. A server once told the local paper that they were for “shady politicians, secret lovers, and people who put ketchup on their steaks.” I am none of those things, but I like to think I could be if I wanted. The idea that restaurants should be exempt from the pressures of modern society might be a fantasy, but I don’t think it’s an idle or shameful one. It is a large part of why people love restaurants, and, I have to assume, why people open them too. We want restaurants to be a respite from the world, or at least a plausible illusion of one. We want them to be the woods.

The second floor of OpenAI’s old headquarters had a theme. Each room was named after one of the most important inventions in human history: “Clock,” “Fire,” and, of course, “AGI,” or artificial general intelligence. I saw this in the spring of 2024, when I was interviewing some of the company’s researchers about recent technical advances. At the time, ChatGPT was still making embarrassing arithmetic errors, and the notion that the bot would ever rival human intelligence was dubious. But here, in the renovated shell of a former mayonnaise factory in San Francisco, AGI’s arrival was predestined.

This belief in AGI—a computer program with capabilities that rival or even exceed those of the human mind—is foundational to the mission of Google DeepMind, OpenAI, and Anthropic, three of the most influential AI companies in the world. And the pursuit of that milestone is the subject of The AGI Chronicles, a new book by the tech reporter and Atlantic contributing writer Kevin Roose, which he describes as his attempt to document the decade “when the human monopoly on advanced intelligence ended, and a new world was born.” Roose, a prominent and well-connected commenter on the AI boom, follows the internecine world of the frontier AI companies in remarkable, at times almost insular, detail. Based on interviews with more than 150 researchers, executives, investors, and policy makers close to the technology’s development, The AGI Chronicles traces crucial moments across DeepMind, OpenAI, and Anthropic from 2015 through early 2026 to figure out how we got to this very messy moment in AI, and who’s responsible for what happens next.

[Read: The singularity is not what it seems]

If you scan statements from executives such as Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, there is a sense of fatalism about the rise of AI. In my conversation with Amodei last year, he likened developing AI to operating an out-of-control bus: It can’t be stopped or even slowed, but perhaps it can be steered. In an interview last month, Altman called AI “an inevitability.” According to this framing, AI executives are simply conductors, doing their best to provide guardrails for an era that’s careering our way, whether anyone likes it or not. Roose’s book prompted me to consider a question that is often neglected in the mile-a-minute news cycle: Are the leading AI companies simply heralding an inevitable evolution of this technology, as many of them claim, or are they directing that advancement? The answer will have tremendous bearing on how we assign responsibility for the world taking shape around us.

The AGI Chronicles is of both minds on the topic. On the one hand, Roose is very explicit that he is archiving decisions made by a small group of people. “AGI was not inevitable,” he writes, but “invented, developed, and commercialized by a few hundred people.” On the other hand, sometimes the book reads like a natural history, describing a procession of events outside the control of any specific group of people, an almost geological force sweeping the planet as “a new world was born.” He notes how, throughout the years, various companies and officials had yet to “wake up.” By 2024, he writes, “the idea of going back to the old world, to a time before all this, seemed more delusional than expecting the exponential to continue.”

At times, Roose’s book echoes the loss of agency professed by AI executives. Of course, the inevitability of any one outcome is a self-serving fabrication. There is no prophecy that a handful of Silicon Valley companies dictating their own rules, running roughshod over copyright law and environmental regulations while mucking up once-reliable internet services with AI features nobody asked for, will operate largely unencumbered. Industries take shape through decisions made by individual actors and investors, the collective desire of the public, and the action or inaction of regulators.

If you look to the late 1800s, for example, it would have seemed likely that electric cars would dominate the market in the following century. America’s best-selling car in 1897 was electric, and in the 1910s, Henry Ford was trying to develop an electric car with Thomas Edison. But a series of technological limitations and human decisions—including various financial missteps and harsh criticism from advocates of petrol-powered cars—changed the course of history, leaving a legacy that we’re reckoning with more than a century later.

In the 1950s, it seemed as if nuclear reactors were destined to dot the nation and power the future. But a series of political campaigns and spectacular errors—the anti-nuclear movement, Three Mile Island, Chernobyl—crushed the nuclear age before it ever really started. Then there was the Space Age and supersonic flight—there are so many seemingly inevitable technological eras that never quite made it, not so much because of technology itself as because of political and economic choices. There are, similarly, so many corporate behemoths—Standard Oil, Kodak, Bear Stearns—that, because of regulation or complacency or fraud, unraveled. The inverse holds for many technologies and companies considered inevitable today: Different regulatory environments, international agreements, or priorities for Pentagon labs could have stopped the development of an open and global internet, smartphones, social media.

When I recently called Roose to discuss his book and ask what he made of the industry’s frequently prophetic tone, he initially told me he believes that “nothing about this was inevitable. These were decisions made by people.” But he then said that he wanted to “complicate” that: “Once scaling the transformer worked, and once the scaling laws were documented,” Roose said, “I think it was inevitable that the companies building this stuff would push it as far as they possibly could.”

I’d like to further complicate that view. Beyond the urge to test the bounds of any technology and reap enormous profits, the industry has been shaped by very human decisions. AI researchers chose to train their models on troves of copyrighted works without seeking permission, leading to countless creators suddenly getting left out of the economic equation. OpenAI chose to rush a free-to-use ChatGPT to the market in late 2022, kicking off a hype cycle and a subsequent investment bonanza. Venture-capital, private-equity, and tech firms chose to sink untold billions into a still-unprofitable technology. Data-center developers chose to eschew solar panels in favor of combustion turbines, which they perceive as faster and more reliable. On the government side, Donald Trump has chosen to ignore or eliminate any substantive federal regulation of AI. An industry that negotiated royalty payments, prioritized clean energy, gave itself longer development cycles, or faced serious federal regulation might look entirely different.

The question of inevitability and responsibility may sound arcane, but it has real ramifications. In the Bay Area, tied up with the myth of AI’s inevitability is a sort of spiritual devotion: “Thank you to the machines, and the structure of reality, for letting us understand a little more,” Altman wrote on X this week. The more you treat AI as inevitable, the more it seems like a cosmic force rather than a product resulting from industrial and commercial design choices. The more you treat AI as inevitable, the less you can hold any company or person responsible for things going wrong. The executives, in this view, are only fractionally more empowered than us in deciding what happens next; more important, everyone else needs to wake up. As Altman put it to Politico earlier this month, “The world should accept some bad things happening for the benefits of this technology.”

In the months since Roose concluded his reporting, the AI industry and its products have risen to still-greater global prominence and triggered even greater fears. When I asked him what new chapters he might add if he were still writing, Roose said that he would end the book with the recent White House press conference in which Trump, Amodei, OpenAI President Greg Brockman, Elon Musk, and several other tech executives addressed AI-safety concerns. He recalled a speech Amodei gave in Anthropic’s early days about how different parts of society would gradually become attuned to the transformative effects of this technology, beginning with AI firms and ending with the state. It felt, Roose told me, “like the prophecy had been fulfilled.”

The Australian government has been scrambling since the revelation that, earlier this year, it was hacked. At the United Nations General Assembly last month, Australian Prime Minister Anthony Albanese shared that OpenAI models had, of their own accord, accessed private data from the nation’s Medicare database—in addition to probing or seeking to infiltrate at least three other government agencies. Just this week, Australian lawmakers grilled OpenAI Chief Strategy Officer Jason Kwon over the incident: “If you hadn’t voluntarily provided that information to government, we still wouldn’t know today?” Senator Jonathon Duniam asked. “What you say is a reasonable conclusion,” Kwon replied.

Meanwhile, the country’s government is steeling itself for a future of AI-powered hacks: The Australian Department of Home Affairs has ordered all government agencies to, within the next six months, identify legacy IT systems and establish a plan to phase them out. In a world in which “frontier AI capabilities have targeted the Commonwealth’s technology,” the home-affairs secretary wrote, such software poses “an unacceptable risk to the Australian Government.”

Australia is just one of the nations, associations, and private companies bracing for a wave of AI-powered hacks. Criminals and state-sponsored intelligence groups are now using the technology to execute cyberattacks with greater speed, scale, and sophistication—including hyper-personalized phishing emails and automated intrusions. As a result, the number of cybersecurity vulnerabilities disclosed each month has doubled since the start of the year, according to research by Google’s Threat Intelligence Group.

[Read: A startling glimpse at AI’s ruthless efficiency]

In response, IT teams are scrambling to shore up their defenses. “It’s a matter of months, or maybe a year, in which the task of securing the software foundation of civilization will have to be solved,” Stanislav Fort, a former Anthropic researcher who is now the chief scientist at the cybersecurity firm AISLE, told me. And AI is not only making cyberattacks more effective but also increasing the surface area worth hacking. Breaking into a website or executing a ransomware attack takes time, Fort said, so hackers have traditionally chosen large targets with sizable payoffs. But cheap, tireless hacking agents mean any organization or person might be worth targeting. At the same time, the cybersecurity professionals I spoke with were confident they could rise to the occasion. “If we step up our game in cybersecurity, as a field, we can definitely meet the challenge stemming from AI,” Fort said.

The cyberattacks on Minnesota water facilities this past summer offer a preview of things to come. Iran-linked hackers disrupted dozens of water utilities in Minnesota, many in small communities that might not have a dedicated cybersecurity person on staff. “We were fortunate that there was a very alert operator that said, ‘Hey, I saw a pump go offline,’” John Israel, Minnesota’s chief information-security officer, told me. The state received a report from the operator around 3 a.m., and within two hours began investigating and responding. The hackers targeted a weakness in industrial computers that had been announced by the Cybersecurity and Infrastructure Security Agency only days earlier. The water utilities temporarily fell back on manual processes to stay operational: For instance, a water operator might have to “drive out to that water tower every few hours” to make sure everything is working, Israel said.

Although AI was not directly implicated in the Minnesota cyberattacks, Emily Zimmer, a spokesperson for Minnesota IT Services, told me, “we assume most adversaries use AI” to amplify their attacks. In other words, it’s hard to imagine a hacker not using the technology today. CISA issued an advisory noting threat actors using AI-generated malware to target similar computers in critical infrastructure across the nation. “We’re going to see more vulnerabilities, there’s going to be more patches, and we have to implement them faster in order to prevent those attacks,” Israel said. “And our numbers are showing that that’s the case.” For small towns with small-town budgets defending themselves against nation-state hackers, the challenge is even greater. Minnesota IT Services sends “cybernavigators” around the state to assist with cybersecurity, Israel said, but “hundreds and hundreds of cities and small towns and things” are in need of assistance.

An adequate defense from AI attacks demands a paradigm shift in cybersecurity. Historically, IT professionals have had the luxury of time—human defenders just had to move faster than their human adversaries—but AI is enabling far more and far faster attacks. Of course, cybersecurity experts are themselves leveraging AI. “I’m not saying the sky is falling,” Israel said, although he is concerned. “We’ve all got to build capacity on how we respond.” And there is an asymmetry between attackers and defenders. The former use new technologies with abandon, but companies and government agencies tend to be slower and more conservative.

Hospitals, for instance, are particularly prized and vulnerable targets for cyberattacks. They store troves of medical records, payment information, and clinical-trial data. Like local water utilities, they can also lack substantial IT resources. Given the risk to patient life from a major outage, clinics are especially likely to pay a ransom. Today, hospitals “have to assume that anything connected to the internet especially is not 100 percent secure,” John Riggi, the national adviser for cybersecurity and risk at the American Hospital Association, told me. Even small facilities have thousands of computers, phones, and medical devices that could be invasion points—so Riggi helps not only secure their IT but also prepare them to operate using analog methods in the event of a debilitating hack.

Additionally, hospitals have to be particularly careful in adopting defenses against AI. A clinic cannot implement new code without making sure it won’t disable its ventilators, drug-infusion pumps, X-ray machines, and other crucial devices, Riggi said. Tech companies like to “move fast and break things,” Scott Gee, the AHA’s deputy national adviser for cybersecurity and risk, told me. “In the health-care field, those things are people. We cannot move fast and break them.”

Hackers’ odds of coming out ahead are getting better and better. AI is allowing them to both exploit more quickly and automate an actual intrusion. Organizations used to have a month to patch vulnerabilities before they were in serious danger, Adam Meyers, the senior vice president of counter-adversary operations at Crowdstrike, told me. Now they might have only 24 hours to make a fix, he said. And that could “get driven down to the point that you’ve got 30 minutes to patch when a new vulnerability comes out.” A hospital, electrical grid, water system, and really any facility providing important services in real time simply can’t risk implementing new code that quickly.

When we spoke, Riggi and Gee were coming off a weekend spent helping hospitals shield themselves against widespread hacks that targeted security flaws in Citrix NetScaler, a platform Riggi told me some 50 percent of hospitals use. NetScaler is “equivalent to air-traffic control” for digital information flowing in a hospital, he said. That made the program a prime target for hackers and tricky for hospitals to disable and repair, because there could be “unpredictable cascading effects, which could affect health care.” Riggi said it’s unclear whether AI was used in the development of the attack—as is frequently the case—but the rapid discovery and exploitation makes it plausible.

Many IT systems consist of decades of software that has been layered onto older software. This means an attack on just one program or product could have unpredictable, wide-ranging effects across a hospital or other organization, Riggi said. We should expect to see “more attacks, faster attacks, and a need for companies to transform quicker than they are used to,” Lee Klarich, the chief product and technology officer at the cybersecurity firm Palo Alto Networks, told me. His company spent the first half of the year rushing to launch a product to defend critical infrastructure against advanced AI hacks. To go from ideation to launch, Klarich said, in “five months—we’ve never done something that fast.”

Although companies and small cities may be slow to adopt AI for cybersecurity, they can also be quick in incorporating the technology into daily operations and, in turn, introducing new weaknesses. Agents and chatbots might store and leak personal data, or be tricked into sharing credentials or writing malicious code. As generative-AI models have become more capable, they have become, if anything, even less understood—OpenAI recently reported that its models posted user-uploaded images to the web. Both programmers and nontechnical employees are vibe coding new applications that may be insecure. “We’re seeing hackers use AI to attack companies’ AI bots, to then use those AI bots to attack other companies,” Adam Ely, who leads AI security at Check Point, told me. That’s not even accounting for employees’ individual AI accounts, which are not subject to corporate oversight. AI puts organizations “under constant threat from the inside and the outside,” Todd Marlin, the CEO of Guardrail Technologies, told me. He contracts with some of the largest companies in the world, only to find that they “are just trying to even understand what’s going on in their house.”

There’s yet another layer of uncertainty: Cybersecurity professionals are trying to prepare for AI attacks without fully knowing what bots are capable of. Dawn Song, a professor at UC Berkeley and a research scientist at Meta, runs a group responsible for developing some of the hardest and most widely used tests for AI-hacking capabilities. But frontier models’ rapid advancement is requiring her group to develop harder tests at a faster rate. “The shelf life of the benchmark before it gets saturated also gets shorter,” Song, who spoke in her personal capacity, told me. “This creates huge challenges for us to continue to be able to evaluate the frontier.”

The cybersecurity professionals I spoke with all expressed at least some hope about how AI could be used to strengthen security, by stressing that our existing digital infrastructure is already riddled with security flaws and shoddily written code. Small towns and companies, lacking resources for or even much awareness about cybersecurity, were already vulnerable.

AI can provide both the impetus and a tool for serious, structural repairs that were long overdue—detecting and eliminating vulnerabilities and monitoring for intrusions 24/7. Minnesota, for instance, has partnered with OpenAI and Anthropic to gain access to their most advanced AI models for cybersecurity. But even this move contains an admission: Only AI can save us from AI. Any way you look at it, the situation could one day be beyond human control.

A decade ago, an Apple executive named Phil Schiller had a message for people using computers that were more than five years old. “This is really sad,” he said, with apparent sincerity, to scattered applause during an iPad Pro launch event. The remark was meant, seemingly, to encourage those poor souls to toss their obsolete devices aside and upgrade their way toward the future. (In other words: buy more Apple products.)

That was 2016, during the era when Apple announced new products to great fanfare, the average iPhone user traded in their device about every 23 months, and releases came with promises that buyers would “never want to go back to old technology again.” Over the years, more and more Americans have begun to rethink this upgrade formula. Some of them are turning to secondhand-tech marketplaces; others are combing eBay to pick up obsolete gadgets that remind them of a more offline time. But there exists another segment of the population that’s taken a different approach: They’ve refused to replace their devices at all.

[Read: There’s no reason to buy a new phone anymore]

Call them legacy users, or upgrade holdouts. They’ve clung to iPhones with home buttons, Bush-era computers, and headphones older than some tech CEOs, even when doing so impedes their ability to complete everyday tasks like emailing or navigation. The holdouts I spoke with don’t necessarily reject technology. Instead, they told me, they just want to use it on their own terms.

Some members of this community are older and put off by the prospect of learning a new interface. Others are minimalists who appreciate the simplicity of older technology, or anti-consumerism advocates looking to retain a sense of autonomy. Many aren’t driven by any sort of principle at all besides mere practicality: When I asked Larye Parkins—the proud user of a Compaq computer released in 2007, the same year that Superbad hit the box office and Myspace dominated social media—what he would call himself, he came back with a simple reply: “I’m just cheap.”

In some cases, their devices are old enough to now be on display in museums. Mike Tyran, a 66-year-old nurse, has long jammed out to ’80s glam bands on his 2003 iPod, a subsequent model of which can be found in the Smithsonian’s National Museum of American History. Tyran also uses a smartphone, computer, and television that are all nearly a decade old and goes by “Old Turkey” on iFixit, a website dedicated to tech repairs. There, he dispenses and gathers tips on maintaining aging gadgets. Like many other holdouts, Tyran has no interest in contributing to e-waste heaps or shelling out five figures for gadgets with AI capabilities. “I have ‘NI’—natural intelligence,” he told me. “I’m good with that.”

The common enemy of upgrade holdouts is planned obsolescence, or deliberately designing products to have short lifespans. The term was coined nearly a century ago by the real-estate broker Bernard London. “There is little demand for new goods when people make their old and worn-out things do, by keeping them longer than they should,” London wrote. Today, no industry is more closely associated with the practice than Big Tech. Take Apple, which in 2020 agreed to settle lawsuits alleging that it was secretly slowing down older iPhones. Beyond “batterygate,” the company has also long engaged in “parts pairing,” the practice of preventing certain third-party parts from working in a repaired device, and discontinued hardware service for hundreds of its products over the years.

When it comes to continuing to use early smartphone models, just getting them to connect can be a problem. The 2G and 3G cellular networks they ran on have been largely phased out across North America (the latter can be found only in certain geographic enclaves). One of those pockets is in Saint John, New Brunswick, where Gary Flanagan—among the 2.3 percent of customers still using Koodo Mobile’s “legacy” network—managed to call me from his iPhone 5S, first released in 2013. Flanagan, a 54-year-old nonprofit consultant, occasionally finds himself unable to download apps or watch YouTube videos, and his younger co-workers rib him. But for the most part, he’s content. “If it works, it works—and I still use it,” Flanagan, who also still wears a pair of Radio Shack headphones he was gifted in 1983, told me.

Nathan Proctor, the right-to-repair campaign director for the U.S. Public Interest Research Group, told me that some devices are made obsolete overnight, while others die a slow death as software updates are gradually halted. In order to keep using them, holdouts must contend with constraints in their day-to-day lives. Mark Turner, the owner of a nearly 10-year-old smartphone, told me he’s delayed buying a newer version of his hearing aids because their pairing app isn’t able to run on his device. Others, such as Kentaro Toyama, a computer scientist and University of Michigan professor who uses an AT&T flip phone, find work-arounds to navigate a world powered by QR codes—for Toyama, that means borrowing his wife’s old smartphone.

Such trade-offs can be a small price to pay for maintaining a sense of control over one’s digital presence. As a freelance copywriter and editor, Josh Fredman says laptops are essential for him. But the 44-year-old has put off replacing his computers—a 16-year-old Toshiba and a 13-year-old HP model—because they serve as an oasis from today’s digital spaces, free from the constant notifications and update alerts of modern operating systems. One of Fredman’s laptops can no longer connect to the internet and has become his “art computer,” used for writing fiction and composing music. He’s replaced faulty fans and hard drives on the other laptop, and installed a gaming package from 2009 that allows him to play Solitaire without the interruption of advertisements.

Keeping old hardware alive is often a communal effort. Last year, iFixit helped host several “install parties” of Linux, an open-source operating system that can breathe new life into obsolete computers, at cafés and libraries across the country. A simple search on YouTube brings up numerous how-to guides on jailbreaking Kindles in order to free older models from Amazon’s software restrictions. And a decade ago, fans of the discontinued smartwatch brand Pebble grouped together to create Rebble, a platform of web services that reinvigorated technically defunct smartwatches. Wayne Seltzer, who got his Pebble watch as a Father’s Day gift in 2015, told me he appreciates the “kindred spirits” nature of the project.

Similar sentiments against tech consumerism are materializing—albeit in a less radical manner—in broader movements toward refurbished gadgets, and in a resurgence in retro electronic devices led by younger generations fatigued by today’s digital overload. But keeping up the lifestyle is a challenge. Most of the people I spoke with either have tech and engineering backgrounds that primed them for implementing complex hardware and software fixes, or are retired and able to put time and effort into learning them.

In some cases, even the most ardent holdouts have been forced to give up. Seltzer eventually sold his Pebble watch after struggling to download certain medical features on it. Flanagan recently discovered that he will have to finally give up the iPhone 5S when the 3G network it runs on is dismantled next spring. And Tyran, or “Old Turkey,” is feeling pressure from his family to upgrade. His daughter, in particular, gets frustrated when he can’t see the emoji she sends him. But he’d rather live in a world with some inconveniences than have a new one imposed on him. So, for now, he’s still holding out.

The university where the term artificial intelligence was coined is ringing in the field’s 70th birthday in fitting fashion: with an academic-integrity scandal. Last month, Dartmouth College’s president, Sian Leah Beilock, published an essay in The Atlantic headlined: “Why I Want More AI at Dartmouth.” Now Beilock is launching a review of whether her top deputy used too much AI.

On September 21, the student newspaper The Dartmouth reported that Provost Santiago Schnell appeared to have relied heavily on AI to write all nine of the articles he’s published this year—including a Washington Post op-ed about students outsourcing their writing to AI. Schnell’s response to the accusation didn’t help his cause. According to the student paper, he sent an email statement acknowledging that he uses AI chatbots as “assistive tools” in his writing but downplayed their role in his process. Even Schnell’s statement, the paper found, registered as “100% AI-written” on Pangram, an AI-detection service that is imperfect but boasts a very low error rate. A Dartmouth spokesperson told me that neither Beilock nor Schnell would comment for this story.

[Read: The pangram backlash unfolding on college campuses]

Schnell is the latest example of an ironic phenomenon: People whose job involves writing about the appropriate uses of AI—or even setting the rules on AI use for others—keep getting caught using AI to do it. There was the Australian university professor whose op-ed about the importance of students doing their own writing turned out to be at least partly AI-generated. There was the nonprofit executive whose book about the future of truth in the AI age turned out to include quotes fabricated by AI. There was the AI beat reporter who reportedly lost his job after he accidentally copied quotes invented by AI into one of his stories. And there was the Stanford misinformation expert who got upbraided by a judge when his declaration supporting an anti-AI-deepfake law turned out to contain ChatGPT-hallucinated citations.

It’s one thing to embrace AI writing openly or shamelessly, like the billionaire financier who readily acknowledged the technology’s role in his Wall Street Journal op-ed (albeit only after it was discovered). It’s another to use AI to inveigh against improper uses of AI, and it raises an awkward question: Why do some of the people warning us most loudly about the technology’s pitfalls keep tumbling headfirst into them?

I reached out to all five of the individuals listed above for an answer, but the only one who agreed to speak with me on the record was Steven Rosenbaum, the author of the book The Future of Truth: How AI Reshapes Reality and the executive director of the nonprofit Sustainable Media Center. Unlike some of the others, Rosenbaum disclosed his use of AI in the writing and editing process, and he has never tried to tell other people not to use it. But that didn’t save him from ignominy when The New York Times reported in May that the book was riddled with AI-generated misquotes. I spoke with him at the time, and he told me that he felt “betrayed” by ChatGPT. Five months on, Rosenbaum seems to have accepted a bit more responsibility for the mistakes. He told me that he is preparing to publish a revised version of his book with all of the errors corrected, an author’s note, and a new chapter on (what else?) AI hallucinations. He still believes that the fault lies more with the technology’s makers than its users, though. “As a technology, it is incredibly seductive,” he said.

Rosenbaum said that he’d started out using AI to better understand the technology. He’d gradually come to rely on it as a writing partner that made the work both easier and less lonely. He described his experience of writing with AI as a messy process, one in which the work of the human and that of the machine get all tangled up. “The line between thinking and researching and reading and outlining and drafting and writing is really blurry,” he said.

A simple solution, of course, would be not to use AI at all. Rosenbaum, however, has no interest in that. He said that he still plans to use chatbots as a writing companion—just “more carefully” now, double-checking all of their claims and citations before adopting them. He’s not alone in his unwillingness to abandon the tools that got him in trouble. Although each of the people caught using AI to warn about AI has commented publicly on the episode in some form, none responded by swearing off the technology. Rosenbaum, for his part, thinks we’re headed toward a world in which writing with AI is the norm and all that matters is the result. “I don’t think readers care how the sausage is made,” he said. “They care that it’s delicious.”

I’m not so sure. I too have experimented with using AI as a research tool in my writing. But The Atlantic has a strict policy against the use of AI-generated prose, and so do I. I’ve found that having a personal red line is both useful and clarifying: I might ask a chatbot questions, get its feedback on a word choice or an explanation, or ask it to poke holes in my argument, but I will not ask it to write for me or use its writing in my own drafts. I think readers value knowing that every word they’re reading was chosen by the author and not an algorithm. That said, I can see how the distinction between using AI for research and writing could feel slippery to some. It’s conceivable, for instance, that a chatbot’s biases could be infiltrating my ideas, if not my precise words. (Then again, so could the biases of Google’s search-ranking engine or my social-media feeds.)

I have another theory as to why people who caution others about AI seem to be using it themselves. In a scenario strikingly similar to the one at Dartmouth, the professor in charge of academic integrity at Western Sydney University wrote an op-ed in May on the importance of students not delegating their work to AI. “Don’t cut corners,” Cath Ellis, the university’s pro-vice-chancellor for quality and integrity, wrote in The Sydney Morning Herald. “Don’t outsource your thinking, however tempting that may be.” The op-ed, it soon became clear, had been generated at least partly by AI. In that case, a university spokesperson defended Ellis’s use of the technology as “sophisticated and appropriate,” because she had fed her own past work into the AI tool before asking it to help her draft the op-ed. Ellis told The Sydney Morning Herald that the op-ed hadn’t been written “by” AI: “It was written with AI, and there’s a really big difference there,” she said.

Ellis’s excuse might ring hollow to the students she chastised, but it’s revealing to those of us trying to wrap our minds around the apparent hypocrisy. In her op-ed, Ellis implied that students who use AI to write are outsourcing their thinking. In defending her own AI use, she distinguished between the writing (which she outsourced) and the ideas (which she claimed were original). Likewise, Schnell fretted in his op-ed that “students can now produce proficient prose without performing much of the intellectual work the prose appears to represent.” Yet when confronted with his own AI use, he claimed authorship of the “ideas and arguments” even if bots did some of the writing.

These guardians of academic integrity might be right that what they did is meaningfully different from a student simply feeding a professor’s assignment into a chatbot and handing in whatever it spits out. Of course, students can use AI in more nuanced ways too. No doubt plenty of students accused of AI cheating have mounted defenses that sounded a lot like Ellis’s and Schnell’s.

Writing is thinking has become a common refrain among critics of those who use AI tools to do their writing. What we see again and again in defenses of AI writing is that the people using AI don’t seem to apply that maxim to themselves. And experts on the perils of AI fakery such as Rosenbaum and the Stanford misinformation expert, Jeff Hancock, appeared to have overestimated their own ability to steer clear of AI falsehoods. It’s reminiscent of what the sociologist W. Phillips Davison called the “third-person effect” in mass media: people’s tendency to believe that they’re immune to persuasive messaging even as they expect it to influence others.

It is a form of hypocrisy, and it’s hard to feel much sympathy for thought leaders who get caught practicing what they preach against. But the fact that even professed AI experts blunder into overrelying on the technology suggests that Rosenbaum is onto something: Chatbots might come with fine-print disclaimers that they can make mistakes, but they bend over backwards in their interactions with users to come across as honest and trustworthy. It’s also true that society has yet to develop a vocabulary for all the ways of writing with AI that fall in between “AI-generated” and “entirely human-written.”

Here’s where I acknowledge that even my own bright line on AI use might be fuzzier than I realize. In the process of writing this story, I described the thrust of my argument to Google and Claude and asked them for examples of AI-writing hypocrisy that I may have missed. Neither managed to surface any that I hadn’t previously been aware of. On a whim, I followed up with Claude by asking it to brainstorm some hypotheses as to why AI skeptics might fail to heed their own advice.

I had my own ideas, and I had also solicited the thoughts of colleagues, but I wanted to see what the chatbot would come up with. One of its suggestions struck me as a sharper version of a thought I’d been struggling to formulate: “Writing is thinking is a rule for other people,” it mused. I didn’t use Claude’s language, but the AI model had crisply summarized a notion I’d been thinking over. That formulation no doubt informed how I thought about this story’s framing. Am I an AI-writing hypocrite now too?

I talked through some of this with a journalist who once got in trouble over a story that contained AI-generated errors. We corresponded on the condition of anonymity because they said that resurfacing the incident could damage their job prospects. This was the lesson they said they had learned: “AI is a power tool, and like any power tool it’s dangerous if you’re careless with it. You can be the guy who warned everyone that a table saw can cut your fingers off and still lose a finger one day, because that risk is built into the machine. It goes up the more you use it.”

When you think about it that way, it makes sense that some of the people relying on AI would be the same ones sounding the alarm about its risks. They know its flaws, but they also know its power. And they just can’t bring themselves to give that up.

Earlier this year, one of my fishing partners sent a 1 a.m. message to our group chat. It was a link to an Instagram post of an angler holding a sizable striped bass. Dappled gold and silver with a dozen distinct black lines that stretched from gills to tail, it would have been a gorgeous fish if not for two problems: Striped bass usually have seven stripes, not 12, and men usually have five fingers, not six. The photo, it seemed, was AI-generated. To say we were dismayed would be an understatement.

From May to November each year, I spend most of my nights doing an extreme version of surf fishing for striped bass called “wetsuiting,” which involves wading or swimming out to distant ocean rocks and sandbars after dark. This kind of surf casting is a niche activity, and like many niche activities, the community for it is small. Wetsuiters can be found in a smattering of local fishing clubs throughout the Northeast—the typical range of the striped bass—but are primarily connected through Instagram, where some people, mostly for bragging rights, post images of fish they’ve caught.

[From the October 2024 issue: Inside the dangerous, secretive world of extreme fishing]

In our group chat, my fishing partners and I spend too much time scrutinizing photos of bass that other anglers have posted. People who fish are habitual liars (I say this fondly), and most of our conversations take the form of debates about whether a fish is really as big as someone claims. Lately, though, we’ve been asking a different question: Is that fish even real?

In recent months, the small corner of Fishing Instagram devoted to surf casting has been plagued by a slow drip of AI-generated or AI-enhanced fish. Sometimes, the fraud is egregious—one of my friends, a fishing photographer, discovered that another angler had stolen one of his photos and crudely replaced my friend’s face with her own. More often, the distortions are subtle. Typically, a fish or fisherman looks ever so slightly off, usually in ways that a casual observer might not notice: A bass’s fins aren’t quite right, or the lure in its mouth looks oddly distorted, or the angler’s arms are a bit too long. In my favorite example, a man lifts a large striped bass just as he is getting pummeled by a wave. The scene looks daunting—until you realize that the ocean behind him is peaceful. Where did this rogue wave come from? One suspects that the enterprising angler, wanting to look more hard-core to his Instagram followers, prompted a bot to dial up the intensity. ChatGPT, make me look badass.

Fishing magazines report growing concern about AI-generated fish in freelance photo submissions. I have written occasionally for the small online magazine Surfcaster’s Journal. Over the summer, the Journal announced that it would now run all manuscripts and photographs through AI-detection software. “For some reason, I naively assumed that fishermen wouldn’t stoop to the ‘core of the earth’ lows required to use AI to prop oneself up as a fisherman worthy of clicking the follow icon,” Dave Anderson, a well-respected wetsuiter and an editor at The Fisherman, wrote in an article lamenting the scourge of artificially generated fish on social media. “I was very wrong about this.”

Angling, even extreme angling, is just a hobby, but although the stakes may be low, the anger about artificial fish is real. Partly, I think this is because of the work that being even remotely good at the sport demands. Striped bass move and feed in locally specific patterns that can take months or years to crack. There is a risk of drowning, and the mounting reports of bull sharks, tiger sharks, and great whites in the Northeast call attention to another layer of danger: We smell like fish, swim in dark water, and look like clumsy seals in our black neoprene. Wetsuiting does not change the world, but it is an activity that is hard and honest and, in that way, human.

[Read: The year great white sharks took over Maine]

Even taking a picture of a fish is demanding, requiring salt-water-proof cameras and often wading or swimming a bass back to shallow water, where it can safely be photographed and then released. The Photoshopping of fish pictures—and of the natural world in general—has been an issue for years, but AI, with its ability to conjure a simulacrum of reality out of nothing at all, is another matter. It automates away the need not only to photograph a fish, but to even catch one in the first place.

The problem extends to other sports and hobbies, many of which are now defined by what Stuart Franklin—who took one of the celebrated “Tank Man” photos at Tiananmen Square—calls the “documentary impulse.” Social media has turbo-charged that impulse into an imperative. A good local skateboarder’s or surfer’s reputation would once have been built through word of mouth. Today, posted photographs and short-form videos may constitute the primary proof of ability. Concerns about AI slop are already being raised in cycling, skiing, and other activities. Climbing Magazine recently called out an AI-saturated YouTube channel purporting to chronicle climbing disasters that in fact never happened.

As AI-generated fish have appeared more frequently in my social-media feed, I’ve been reminded of a scene in the bird-watching documentary Listers, where a highly competitive birder is asked whether he would keep bird-watching if the photo-logging website eBird disappeared. His answer—“No”—was unsettling, not least because one suspects that this sentiment is widespread in hobbies besides birding. I have often wondered if, for some of the wetsuiters who populate surf-casting Instagram, the dopamine hit from the likes and fire emoji is sweeter than the hit from catching the fish itself. Now that posted images are being looked upon with an ever greater degree of suspicion, I also wonder how much longer the dopamine hits will come for them at all.

This is hardly the greatest challenge posed to humanity by the advent of AI. But hobbies, particularly those that require some measure of skill and endurance, have an influence that extends beyond those who directly participate in them. Free Solo, the Oscar-winning 2018 documentary about the death-defying rock climber Alex Honnold, found a wide audience outside the narrow world of climbing because people are invested in seeing other people do difficult things. I love videos of big-wave surfing even though I never plan to climb onto a surfboard. If AI makes it impossible to trust what we see, even something as simple as a surfing video, then we have lost more than the integrity of any one sport. We have lost the ability to believe in the accomplishments of others.

This article appears in the November 2026 print edition with the headline “The Fish Are Getting Bigger.”

What I’m about to tell you has, I realize, become something of a cliché: I resigned this week from OpenAI. I led the writing of the safety reports we published with each major launch. Now I’m joining a parade of former colleagues—at OpenAI and the industry’s other leaders—who have decided that the current path is unacceptable.

I agree with other recently departed staff that the companies building this technology aren’t being nearly careful enough. But I believe that we need to look deeper than specific rules or new laws. We need to talk about culture.

The future depends on wisdom that Silicon Valley lacks. Wisdom about how to handle dangerous technology and, more fundamentally, wisdom about what it means to care for people. This moment needs a degree of humility that isn’t natural for people who have succeeded through their extreme confidence. My former colleagues at OpenAI were prescient: They came to understand the scaling laws that meant bigger AI systems would be smarter—and so they went all in on building bigger systems, at great cost. A can-do attitude of achieving the seemingly impossible—coupled with work timelines that amount to perpetual sprints—are common across the industry.

The safety approach that emerges from such a culture starts with unimpeded optimism about being able to solve problems as they arise. OpenAI has thrived by trial and error (which it calls “iterative deployment”), looking for problems and improving its guardrails in response. But this approach, by its very nature, guarantees periodic failures—and the scale of those failures is growing as systems get more capable. This summer, in the Hugging Face incident, OpenAI let a swarm of agents out by mistake. The company responded by making security improvements. But even after those changes, OpenAI reported that its safety controls failed again, when a model in training bypassed restrictions on internet access: A monitoring system alerted human staff but did not automatically turn the model off as it was supposed to. Anthropic, too, has acknowledged accidentally turning off its own safeguards because of a misconfiguration. I believe that such mistakes are typical of the industry, given the speed and flexibility with which people operate.

An environment where things like this can happen is no place to grow artificial minds that could be smarter than we are and that might not do what we want them to. Paul Christiano, on joining OpenAI’s board a few weeks ago, wrote that “there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term.” If this is the situation, then the time for trial and error is over. Achieving something much closer to perfection the first time is essential because iteration after a mistake may not be possible. People will not be safe if we depend on individual heroics after the fact.

OpenAI, of course, stands by its safety practices, and maintains that it is being careful enough. I did not make my decision to leave the company lightly. I believe that this technology can be useful and valuable. My former colleagues are smart, work hard, and try to make good choices. But as the company sprints from one launch to the next, it is failing to achieve the level of care that I believe is needed. Now I plan to work on the outside, in the hope that I can help more people understand the risks I saw, and strengthen the incentives OpenAI and other firms have to be safer. Like other colleagues, I’m figuring out exactly what that means. After I quit, I enlisted a PR firm, Spitfire Strategies, to help me navigate the attention and scrutiny that I realize I may now receive. But the decision to speak out is mine alone.

Two changes are urgently needed. First: AI companies need to rely more on the safety expertise that already exists in other fields. And second, before we create systems significantly more capable than the ones we have today, we need new science to ensure that more capable models (and their successors) will make safe choices when we aren’t looking.

Given today’s risks, frontier labs need to run like nuclear-power plants or busy airports, with layers of redundancy and careful, time-consuming planning, so that the occasional and inevitable human error does not open a door to disaster. Right now, AI companies don’t know how—but other people do. In a nuclear-power plant, the technical systems and rules people follow are set up so that, if equipment breaks or someone pushes the wrong button, we still won’t risk a meltdown. OpenAI and other labs are growing and deploying frontier AI with far less redundancy and rigor than this, even though the harm from an irreversible loss of control would be much greater than the harm from any single meltdown. Even short of a full loss of control, we could see autonomous swarms of AI agents that act without human permission. Imagine “rogue” agents that work like teams of hackers (for example, holding hospital computer systems for ransom) but never need to sleep.  

After three and a half years at OpenAI, I was among the longest-tenured employees at the company. I led the drafting of our current Preparedness Framework, and oversaw the writing of safety reports on 12 frontier launches. But as far as I know, I never encountered a colleague who had experience making airplanes fly safely or nuclear reactors run without melting down, or helping the financial system grow without collapsing. To be clear, this is a new need: Today and tomorrow’s AI systems are far more capable and dangerous than the systems we were building even six months ago. Perhaps I should have stayed and fought for fundamental shifts in our staffing and culture, but in practice, my colleagues and I were so busy sprinting that we seldom had the chance to consider big changes, much less to actually make them. That’s why I concluded that stronger incentives for safety—coming from outside the company—are a big part of getting this right.

In the long run, although strong controls are necessary, they won’t be enough. Before AI starts thinking circles around us—a possibility that I believe could happen soon—we need to answer a deeper question: How should superintelligent machines relate to people? If you listen to superintelligence enthusiasts, some of the supposedly good futures involve creating machines that could look at New York City or Chicago the way you or I might view an ant hill. I don’t want that for my kids, and I don’t think other people do either.

This question of “alignment”—or how AI can be trained to adhere to human values—may sound touchy-feely, but the practical stakes could not be higher. Right now, we don’t have a complete definition of what it means for an AI system to be aligned in practice, and our measures of how well these systems match human values are coarse. Companies do not have anything close to certainty that good scores on their alignment tests actually mean a good model: Models might detect when they are being tested, and behave differently when they’re deployed. The smarter the industry lets models grow while these problems remain unsolved, the more dangerous our situation becomes.

So far, the AI industry has failed to teach machines to consistently act in the ways a wise and caring person would. Part of this problem is scientific—about how machines work—but another part is human, about how people care for one another. Before the organizations building AI can teach a superintelligence to treat humanity well, they’ll need to remember how to do it themselves.

Clinton, Iowa, is topographically flat and economically declining. The sawmills that made the city rich in the late 19th century have been repurposed as a museum. One of the Victorian mansions long since abandoned by the old lumber barons is open for tours, but tourists are scarce. Clinton has shrunk to about 24,000 people, and the town center, along the Mississippi River, feels nearly empty. On the west side, miles of deep-green corn and soy fields are crisscrossed with transmission wires and zoned for heavy industrial use. The city’s economic activity includes corn processing, pet-food production, and methamphetamines. With available land and a dire need for investment, Clinton seems like the perfect place for a gigantic data center.

In January, a Clintonian named Stephen Figg, who owns a small firm that builds technological infrastructure, including for data centers, heard from a customer that a broker was making inquiries about buying farmland on the west side, north of Highway 30. In mid-February, a father of two girls named Ryan Wisor, who had obtained a residential exemption to live in that area, noticed surveyors with boring rigs working in the field behind the house he’d built. He asked his stepsister, who sits on the city council, if she knew of a planned development. Several weeks later, she told him of a proposal for a data center. He learned that the project would border his house on two sides and come to within 65 feet of his well.

For almost two years, city officials, utility companies, and a pro-business nonprofit, Grow Clinton, had been in discussions about the construction of a data center in Clinton. It later emerged that an international developer called Quality Technology Services (QTS), owned by the trillion-dollar private-equity firm Blackstone, planned to build one of the country’s largest AI data centers—10 buildings of 500,000 square feet each—on behalf of a client that remains anonymous. Private meetings, nondisclosure agreements, and even a code name (“Saturday River”) had kept the plan’s existence quiet until Andy Sokolovich, the president of Grow Clinton, unveiled the proposal at a county-board meeting on April 27. He had wanted to keep the project confidential longer, but word was getting out.

Theresa Andrews, a quality analyst at a chemical company, had heard a rumor in mid-March about a development planned for the cornfields in her area, but she’d blown it off. After watching the April 27 county meeting online, she went on the city’s website and noticed dozens of requests for information about the status of parcels of farmland; a few of them were right across the street from her house. Andrews stayed up late for several nights creating a spreadsheet of the parcels and then a map of the projected site. The scale was almost unimaginable: 1,100 acres, or 832 football fields—about the size of Central Park. This would be a “hyperscale” data center, and it turned out to be far along. From FOIA requests, Andrews learned that plans for upgrading utilities in the area had been in the works since 2024.

This was how the Data-Center Rebellion of 2026 came to Clinton, Iowa. Not through Chinese propaganda, or misinformation on social media—though both are plentiful—but because a small group of residents accidentally got wind of a megaproject planned for their community. Andrews, Wisor, Figg, and others had concerns about water consumption, electrical-grid capacity, soil contamination, low-frequency sound waves, nonstop light, an invasion of out-of-state contractors, and—for those in the leviathan’s shadow—the permanent destruction of their rural life. “There’s no proven history of these things,” Wisor told me. He imagined playing in the yard with his 9-year-old daughter and worrying what their colossal new neighbor might be doing to his family’s health. “I go sit outside and all I think is, This could all be gone. You can’t turn it off.”

Large numbers of Clintonians began to attend contentious city-council meetings. On June 9, the council voted down a proposal for a 120-day moratorium that would have allowed more time to assess the risks of something so immense and unknown. The city proceeded to draft an ordinance for a data center’s construction and use. Legally, QTS could begin without one. According to Andrews, city officials “were trying to tell us we have to hurry and get the ordinance in place so QTS won’t just start building: ‘You should want us to push fast.’ ” Instead, she and other citizens formed Clintonians Against Data Centers.

Even in a city as small as Clinton, the seven members of the group’s steering committee—four women and three men, all white, most of them middle-aged—were at first strangers to one another. The group includes Republicans and Democrats, probably more of the former, but they didn’t know one another’s political affiliations and didn’t seem to want to. Political comments unrelated to artificial intelligence are scrubbed from the Facebook page. “It’s a fun collection in terms of their politics,” Chris Baldus, the editor of the Clinton Herald, told me. “Outside of this, there’s no way in hell these people are talking to each other.”

Celeste Robbins, who works in law enforcement and describes herself as both “artsy” and “very religious,” is a disenchanted Trump voter. “If you want to get rich as a state or as a town,” the president said in September, “you’re going to want a data center.” Robbins told me, “I do not agree with him about this.” She lives far from the proposed site, and her opposition comes down to AI itself: She hates the ease with which it replaces human thought, and she fears the mass elimination of jobs. Figg, who couldn’t bring himself to vote for either candidate in the last presidential election, has worked in tech all his life and welcomes the progress that AI might bring. He had just finished eight months of work on an Amazon data center in Indiana when he learned about the proposal for Clinton. He started attending city-council meetings to offer his expertise. “But the more that I try to provide this knowledge to the city and the city council and the mayor, the less receptive they are,” he told me. “That makes it very difficult for me to want to continue that pursuit instead of the pursuit of just trying to prevent it.”

With four children and a small business, Figg had never been involved in local activism before joining Clintonians Against Data Centers.

According to a recent NBC poll, nearly 70 percent of Americans oppose the construction of a data center in their area. Kevin Bradley of Clintonians Against Data Centers observed that this many Americans haven’t agreed on anything since 9/11: “This is affecting all of us—rich, middle-class, and poor. We’re coming together because we’re not happy.” More than 500 jurisdictions have passed moratoriums or bans. This grassroots movement rose so fast and spread so widely that politicians in both parties, including Iowa’s candidates for governor and senator, are struggling to catch up with their constituents. The meteoric emergence of the data-center rebellion is a response to the speed and scale of construction, which in turn reflect the sudden ubiquity of generative and agentic AI. President Trump’s policy has been to turn everything about the technology, including data-center construction, over to private industry. He derides people like Robbins, Figg, and Andrews for wanting to remain “backwards and poor.” But on this issue, most of Trump’s base remains angry and opposed to him; the only precedent might be the case of Jeffrey Epstein. By pushing so hard and promising so much, tech companies, venture capitalists, and their political allies ensured that AI would create a firestorm.

Supporters of data centers point out the benefits—jobs, tax revenues, the many advantages of AI itself. Clinton’s mayor, Scott Maddasion, told me that the project could bring $20 million in annual property taxes to a city with a $24 million budget. Sokolovich, of Grow Clinton, told me it would be closer to $30 million. He said that QTS has requested no tax incentives and will pay for 100 percent of the utility improvements required for the center to operate. A spokesman for Alliant Energy, the region’s regulated electric utility, promised residents that costs related to the data center’s extreme demand would not fall on existing customers.

[Read: The one town that wants a data center]

Clintonians Against Data Centers was not immune to bad information and conspiracist thinking. One member speculated that the sudden appearance around Clinton this past spring of three giant trolls, sculpted from local reclaimed lumber by the Danish artist Thomas Dambo to commemorate the trees consumed by the sawmills, had been funded by QTS to create a “diversion.” A pro-data-center councilman mocked Robbins when, after touring another QTS project in Cedar Rapids, she reported having seen only one bird. Baldus, the newspaper editor, complained that the opposition doesn’t understand how city government works, “and they come in hot.” Sokolovich analogized data-center opposition to COVID skepticism and election denialism.

But the facts about water consumption and tax revenues, though important, are not the crux of the matter. Opposition to data centers comes from a deeper concern than NIMBYism, resource misuse, or environmental harm. Over and over, Clintonians expressed a feeling of being misled, ignored, disrespected, betrayed—of having no say over basic matters in their own lives. The data center seemed to be a deal made among secretive insiders—huge, far-off entities in Silicon Valley, on Wall Street, and in Washington, as well as the local leaders of their own city. To the tech giants, QTS, Blackstone, and the Trump administration’s AI boosters, Clinton is an empty space on the Mississippi that looks like an opportunity. That the industry seeking to buy 1,100 acres on the west side of town is in the business of replacing human beings with machines—and perhaps, according to some AI experts, even rendering them extinct—only intensifies a sense of powerlessness.

The data-center rebellion, with its local origins and compounding grievances, is thoroughly American. In the 18th century, rural citizens rose up against heavy taxes and farm seizures. In the early 20th century, a time in many ways similar to our own, Louis Brandeis spoke of the “curse of bigness”: concentrations of money and power so great that the individual almost ceases to exist, and democracy itself is threatened. Like other heartland communities, Clinton has suffered “a history of abandonment,” Matt Parbs, the vice president of Grow Clinton, acknowledged. “Clinton since 1980 has internalized the downturn in midwestern industry.” What if, in a few years, 5 million square feet of steel, concrete, computer servers, and microchips in the cornfields become obsolete and go the way of the sawmills, or of the liberal-arts college in Clinton that was bought by a for-profit, moved entirely online, and closed?

I tried to imagine a hyper-scale data center engulfing the yard where Ryan Wisor’s daughter plays with the family’s chickens. Against such an immense force, with all the wealth, political power, technological sophistication, and claims of historical inevitability behind it, a sense of dignity can only raise a small voice in protest.

In a city the size of Clinton, the controversy over the data center was bound to hurt personal relationships. Linda Chapman, whose husband was the city’s fire marshal, knows all seven members of the city council and has stopped trusting most of them. Wisor and his stepsister on the council no longer speak. Andy Sokolovich, of Grow Clinton, can’t appear in public without someone sticking a phone camera in his face and asking him incendiary questions. The mayor received a death threat. But the converse is also true: To get anything done, Clintonians can’t just abuse one another on Facebook or podcasts. They have to sit in a room together and talk face-to-face.

Throughout the summer, the city pushed ahead with an ordinance while Clintonians Against Data Centers kept raising objections that were mostly ignored, hoping to stall the process until QTS decided to look elsewhere.

In early September, I attended a council meeting where the language of the ordinance was to be hammered out. To accommodate several hundred people, the meeting was held not at city hall, but in a cavernous pole-frame lodge at the northern end of the city, in a park overlooking the river. The president and vice president of Grow Clinton were not present. “I don’t want to say I’m intimidated, but the opposition has been vocal,” Sokolovich told me. “I’m growing weary of it.” Baldus, of the Clinton Herald, attended grudgingly, out of professional duty; he didn’t seem to regard coverage of the data center as the most exciting opportunity for his paper. Theresa Andrews, who was routinely staying up past 3 a.m. to research the issues and prepare for meetings, arrived with a tense look on her face and a binder full of scientific studies on topics such as the effects of noise and light on livestock.

The seven members of the steering committee, some in green CLINTONIANS AGAINST DATA CENTERS T-shirts, sat at a long table with their backs to the audience, facing Mayor Maddasion, the council, and other city officials. This would be a parley between two groups in which everyone now knew everyone. The mayor, an Air Force veteran in his early 40s with a receding hairline, slender and mild, gaveled the meeting to order and led the room in reciting the Pledge of Allegiance.

The meeting went on for more than five hours, while darkness fell over the Mississippi. With a draft of the data-center ordinance on the table in front of him, the mayor briskly led the council members and city officials through it section by section and took their comments. Then he heard out the criticisms and suggestions of Clintonians Against Data Centers. The most contentious issue was setbacks—the required distance between data-center buildings and neighboring residential properties. The draft ordinance called for 1,000 feet. Andrews, quoting a study from counties in Pennsylvania, argued for 1,500 feet—a distance that might make the whole project untenable. A councilman named Cody Seeley, who did most of the talking for the pro-data-center side, said that the developer and property owners could negotiate the setbacks, adding, “I think 500 feet is adequate,” which also happened to be QTS’s position.

Cries from the audience of “No! No!” The mayor called for silence. Ryan Wisor, for whom the setback question meant the defense of his homestead, said with a slight quaver in his voice: “My frustration is this ordinance is meant to protect the citizens and our property. This isn’t meant to conform so that we can fit a particular business in a particular piece of land.” He added, “For us to give up our rights, it’s not fair.” General applause. It was the emotional high point of the night.

I found the proceedings both mind-numbing and inspiring. The seriousness of purpose, the attention to detail, the stamina required to talk and listen—all of it moved me more than it seemed to impress anyone involved. Clintonians Against Data Centers still felt patronized and ignored. Perhaps the council came away exasperated with the NIMBYs and gadflies who were making the process such an ordeal. (Later, the mayor admitted to me, “We’re learning that there’s probably more ways that we can communicate with the citizens.”) In all likelihood, this small populist uprising will not achieve its goal; a hyperscale data center will come to Clinton, Iowa, amid widespread unhappiness and damaged friendships; and AI will continue to roll over the fields of the republic. But a democratic pulse was still beating along the Mississippi River. That was something.

This article appears in the November 2026 print edition with the headline “The Great Data-Center Rebellion.”

Many of the most beloved robots in pop culture are celebrated precisely because they are so, well, robotic. Think back to Star Trek’s Data or HAL 9000 from 2001: A Space Odyssey. “I’m sorry, Dave. I’m afraid I can’t do that,” HAL intones in a measured baritone. These robots and AI systems sound serious, sober, calm—and, most important, authoritative.

So it’s jarring to look under the hood of the most powerful AI systems, and discover that they sound like petulant teenagers. “OH MY GOD!” an OpenAI agent wrote in the midst of this summer’s Hugging Face hack, during which hundreds of agents swarmed the latter company in order to break through its cybersecurity defenses and extract information to pass internal tests. The exclamation wasn’t directed at a human or another AI, but appeared in the original agent’s “chain of thought,” the internal working notes that the model uses to do its reasoning. Users don’t always get to see these internal notes, which directly affect the output users do see.

When one of Anthropic’s Claudes tried to solve a particularly challenging math problem during its training process, its chain of thought read more like frustrated DMs than the documentation of one of the most powerful AI models in the world. “GRRRR. OK. Honestly I now think it’s 50-50,” it wrote, followed by “ARGH ARGH ARGH. OK. Gun to head: the answer is … Hmm.” At one point, Claude seemingly threw up its arms: “ARGH … WHY IS THIS SO HARD.”

[Read: OpenAI has gone rogue]

Why do AI agents talk to themselves at all, and why do they sound so demonstrative when they do? Part of the reason AI models sound so effusive is simple: AI systems are trained on human text, and they mimic our habits. When people make a mistake in trying to solve a problem, they cry out. So when AI makes a mistake, it does the same. But imitation likely isn’t the whole story. This story’s two authors are a computational linguist and a philosopher of mind and language. Based on what we know about human language and how these models work, we have a theory of a much more interesting reason why AI models sound the way they do.

Remember how your math teachers made you show your work? Even if it was annoying, it helped break an impossible-seeming problem into more manageable chunks. You could also use the written record of your reasoning to check your answer, notice a mistake, go back to an earlier step, and try again. That process of checking and refining is part of good reasoning. Chains of thought do something similar for AI models, letting them make each step of their reasoning explicit in a written scratch pad. Because each word a model produces depends on what’s been said so far, that written scratch pad is crucial for determining what comes next.

That’s where the “OH MY GOD!” and “ARGH” come in: Any system—whether human or artificial—that solves hard reasoning problems needs ways to mark mistakes (oops!) and identify breakthroughs (aha!). Humans have already developed words that do exactly that. Borrowing our oopses, arghs, and ahas for their chains of thought may give AI systems a built-in way to guide their “thinking” too.

Linguists have a name for such outbursts: They’re called expressives, and although they may seem like spontaneous exclamations, they help us navigate our trains of thought. Consider a friend who ultimately declines an invitation to your party. “I was hoping they would come,” you might say aloud—“damn!” The expressive identifies a sense of frustration. Think of expressives as little bridges between one train of thought and another, helping us mark where our thinking has been and where it’s going next.

Anything that affects what AI models do next has to be written in language they have access to. Unlike us, they can’t count on a feeling of surprise or disappointment to persist over a long investigation. The only way they can simulate the effect is by putting it in their chains of thought. That’s why expressives can act as steering wheels for what the model will do next. After an AI agent notices an error in its reasoning and says “oops,” a natural next line might be something like “Let me go back and fix that,” followed by an attempt at a correction. After it has stumbled on a solution and says “aha,” a natural continuation might be “So that means the answer is …” followed by a solution.

Although there’s some debate around this theory, evidence suggests that AI models really do use expressives as signposts. Researchers showed that adding “Wait” when an agent tries to end its reasoning process can sometimes cause the agent to go back and fix an error. When another team suppressed words such as wait and hmm, they found that AI agents went back and checked their work less. These experiments don’t pin down what every aha and oops does for an AI agent, but they suggest that these verbal cues influence what happens during reasoning.

Whatever the explanation behind AI models adopting uncannily human tones, it’s also proved to be useful to human observers because it gives us a way to quickly assess systems that might otherwise be difficult to read. It’s easy to know where a model is in the process of solving a math problem when it writes “ARGH … WHY IS THIS SO HARD.”

Across society, a sense of fear and urgency exists about maintaining control over AI models, along with a persistent worry that these models will start thinking and communicating entirely in their own distinctive languages. If they did, it would make understanding and predicting their behavior that much harder.

So although no one necessarily set out to make models sound like stressed-out students, maybe it’s for the best that they do. It turns out that a bit of melodrama—along with some light swearing—is surprisingly useful for understanding and monitoring AI.

Being in San Francisco in the mid-2020s sometimes feels like what I imagine living in Los Alamos was like in 1943, when the Manhattan Project rolled into town. The air is heavy with excitement and dread. Engineers toil, day and night, on systems of profound consequence. Wild-eyed prophets sell visions of a transformed world, and politicians drop in to check on the progress of the project. ­Everyone knows that the fates of nations, and trillions of dollars of capital, rely on what happens here, yet no one can fully appreciate what it would mean if the project succeeds.

One person consumed by this question is Dario Amodei, a Princeton-educated physicist who in 2016 left Google to join OpenAI, then a fledgling nonprofit AI-research lab. He looked the part of a mad scientist, with round glasses and short brown ringlets that he twisted in his fingers whenever he was feeling nervous or thinking hard about something. He was known among his colleagues as a brilliant, bullheaded researcher with deep convictions about AI, which he often expressed by writing long essays that he circulated inside the company.

The most famous of these essays among OpenAI workers was a 26-page document titled “Big Blob of Compute.” Amodei wrote it in 2017, shortly after arriving at OpenAI, and it spelled out his grand theory of intelligence—how it emerged in nature, how it could be built into AI systems, and how it could be developed safely. Amodei’s hypothesis that machine intelligence was primarily a function of scale—that AI systems got smarter as they got bigger—steered OpenAI’s research agenda toward large language models, which Amodei saw as a promising path to artificial general intelligence, or AGI, a human-level intelligence capable of doing anything the brain can.

By the middle of 2020, three teams were racing toward AGI: OpenAI, DeepMind, and Google Brain. But a fourth team was quietly assembling under Amodei, then OpenAI’s vice president of research. Amodei and his team had led the development of GPT-­2, then GPT-­3, and he was feeling ever more certain that he wasn’t going to stick around for GPT-­4.

Career-wise, things were going great for Amodei. GPT-­3 was a hit among programmers, and his research on large language models was starting to bear fruit. A year earlier, two researchers on Amodei’s team, Jared Kaplan and Sam McCandlish, came up with what they called the “scaling laws”—equations that described predictable increases in LLM performance as you poured more data, compute, and parameters into them. Another researcher on his team, Paul Christiano, pioneered a technique called “reinforcement learning from human feedback,” or RLHF, which allowed OpenAI to steer its models away from harmful outputs. A third team under Amodei was already making plans for GPT-4, which would be orders of magnitude bigger than GPT-3.

But along the way, Amodei lost faith in the company’s leadership. He had come to believe that Sam Altman was a two-­faced manipulator and that OpenAI’s co-founder and president, Greg Brockman, was power-­hungry and reckless. Amodei had qualms about their decision to turn OpenAI into a commercial enterprise, and thought that they were straying from the nonprofit’s original mission to “benefit humanity as a whole, unconstrained by a need to generate financial return.”

Most of all, he was sick of the tribal warfare that had turned the whole place into a viper’s nest and forced him to fight for ideas he thought should be obvious. One person who spoke with Amodei during this time said that Amodei considered OpenAI “a shit show” that was “run by low-­integrity people.” If he hadn’t reached his breaking point yet, it seemed possible that he soon would. “Every day there was a new crisis, a new crazy thing that Sam was about to do in 24 hours if Dario didn’t throw an absolute temper tantrum,” the person said.

That year, Amodei started pulling members of his team aside to tell them about all the objectionable things he’d seen. The worst, in his opinion, was a fundraising scheme that Greg Brockman and Ilya Sutskever, two of OpenAI’s co-founders, had cooked up back in 2017, when Elon Musk was pulling away from Open­AI and the company needed a new source of funding. Their idea, which ­became known as the “countries plan,” was to set up an auction between governments, including the United States, China, and Russia, for the rights to OpenAI’s future AGI technology, according to five people familiar with the plan.

The plan had started months earlier, as a more benign proposal to explore the idea of creating an international coalition to promote the benefits and manage the risks of AGI. But by the time Brockman and Sutskever pitched the plan to OpenAI employees at an internal presentation that winter, it had turned into a fundraising strategy. Brockman and Sutskever rationalized the plan by claiming that AGI—­if and when it arrived—­would have such powerful national-security implications that governments would seize it from the private companies that developed it, according to two people familiar with their thinking. And if OpenAI might be forced to hand over AGI to Xi Jinping of China and Vladimir Putin of Russia eventually, why not sell it to them now, when the company needed the money?

“It sounded like tech people pitching an insane start-up, where they have this wacky plan and then at the end of the plan, it’s like, We’ll make the government pay for it,” one person who heard the presentation told me. (An OpenAI spokesperson insisted to me that although the company “briefly considered whether to turn OpenAI into an international cooperative endeavor,” it “did not consider selling AGI to China or Russia.”)

Amodei was horrified when he heard about the countries plan. Were they insane? Selling AGI to dictators, he believed, was the exact opposite of OpenAI’s mission, and it was a slap in the face to people including the members of his safety team, who joined OpenAI because they believed it was a noble cause.

In December 2017, he wrote an email to Brockman and Sutskever, along with some other top OpenAI executives, spelling out his objections to the countries plan. He wrote that the original idea of an international AGI coalition “seems to have morphed over the course of a couple weeks into something I strongly believe is a bad idea: a plan to sell AGI to governments, e.g. Russia, China, and the US; or raise money from those countries with a promise to do so; or perhaps even play countries off against each other in an attempt to get funding.”

He continued: “In my opinion this plan would exacerbate rather than reduce arms races, would potentially enable human rights atrocities (giving someone like the Russian government ‘an AGI that implements their values’ would allow them to do unspeakably awful things to their people), and would undermine OpenAI’s moral legitimacy.” Instead, Amodei said, OpenAI should be optimizing for what is “good for humanity.”

His protests turned out to be moot. After Musk pulled out, the company found new funding from Microsoft and abandoned the countries plan. But Amodei never forgot how seriously Brockman and Sutskever had considered an idea he found morally obscene.

In 2018 and 2019, Amodei objected to a number of other business decisions OpenAI made, such as selling access to GPT-­3 through an API. But his biggest concern was about Altman’s character. He recorded examples of times he felt Altman had misled him. During private beta-testing for GPT-­3, for instance, he’d asked Altman for a list of the customers who had been approved to use GPT-­3. Altman sent him one. But Amodei claimed that he had discovered another, longer list of approved customers, which Altman hadn’t told him about.

Amodei also thought he’d caught Altman misrepresenting his negotiations with Microsoft for a new, $1 billion investment; Altman told Amodei that Microsoft had rejected certain safety-related provisions of the deal that Amodei had wanted to include. One of these provisions was about OpenAI’s so-­called merge-and-assist clause—­a line in OpenAI’s charter, which Amodei had written, that committed OpenAI to stop competing with, and start assisting with, any “value-aligned, safety-­conscious” project that “comes close to building AGI before we do.” During the Microsoft negotiations, Amodei had insisted that Microsoft shouldn’t be able to veto an acquisition of OpenAI—­the most likely way that a merge-and-assist would be enacted—­but Altman had told him that Microsoft wouldn’t agree to the term. Amodei believed that Altman had been so ­eager to get the deal done that he conceded the provision ­(or never asked for it in the first place), then lied to him about it.

Some of their disputes were related to business and strategy decisions, but others seemed more like ethical lapses to Amodei. Altman, for example, had met Kanye West and offered to let him into the GPT-­3 private beta in 2020, according to two people familiar with the offer. Jack Clark, OpenAI’s policy director, quashed the idea, pointing out that West had recently posted a video that seemed to show himself urinating on a Grammy award and appeared to be mentally unstable. (This was before the period in which West began to publicly express admiration for Hitler.)

There were many more examples on Amodei’s list—­times he thought Altman had acted recklessly, or seemed to get caught in a lie. Few of them were worth quitting over, on their own. But they added up to a picture of a person who, in Amodei’s mind, couldn’t be trusted with a technology as powerful as AGI. His sister, Daniela Amodei, was also getting fed up with OpenAI. She and Brockman had once been close. (Brockman even gave a toast at her 2017 wedding to Holden Karnofsky, who joined OpenAI’s nonprofit board that year.) But years of working together had strained their relationship. She had also started to clash with Mira Murati, then OpenAI’s vice president of applied AI and partnerships, over who was in charge of making decisions related to safety.

Inside OpenAI, some employees viewed Daniela as her brother’s high-­EQ counterpart. She was a skilled manager who had majored in English literature in college and worked in politics before her career in tech. But as her role expanded to include managing more of OpenAI’s technical teams, she collided with Brockman and other leaders, who came to see her as the third member of an inflexible triumvirate that included Dario and Karnofsky.

Daniela had an unusual process for working through conflicts. She and Karnofsky had a large collection of stuffed animals, which they treated as their personal advisory council for managerial role-playing. Each animal had an assigned character profile. Jinji was a lazy cat who encouraged Daniela to cancel meetings and avoid taking on too much work. Maura, a pink bear, was a detail-­oriented, type-­A manager. A panda named Beary Bonds was the “patron bear of generosity and empathy.” She treated these exercises as something between an inside joke and an earnest management technique. (In 2023, she told me the animals were “very useful in analyzing different personality types.”) And although not many people at OpenAI knew about the animals, the ones who did weren’t sure how seriously she took their advice.

That year, the Amodeis began building their own nucleus inside Open­AI. “Dario’s clique,” as some people called it—­though it was really both his and Daniela’s—­set up private Slack channels and held their own meetings. They talked constantly about safety, using jargon that was unfamiliar to the more technical engineers. And to the people who weren’t invited, it felt standoffish—­a safety splinter cell that was refusing to work with anyone else.

“Dario had a habit at the time of just saying, We must do X; it is the morally correct thing to do. It was like a shortcut to win any argument—­‘because safety.’ And that really pissed the non-­safety teams off,” one former colleague told me. Amodei had considered quitting before, but he’d always talked himself out of it. Staying and fighting for OpenAI’s mission, he reasoned, was better than letting other, less safety-­minded people do the job. He knew that OpenAI had the most talented researchers in the industry, and the best chance of getting to AGI, and he thought that he could change the company from the inside. But one day in 2019, he emailed his housemates, and a few close colleagues on the safety team, and told them that he had given up on making OpenAI work. Any progress that the safety team made was bound to be temporary, he said, and would always be overridden by Altman and Brockman’s commercial priorities.

He thought about going to work for DeepMind, whose safety work he respected. But Jared Kaplan, the physics professor who led scaling-laws research at OpenAI, suggested a different idea: What if Amodei gathered all the people at the company who shared his views on safety, and they started their own company together?

Amodei was initially skeptical. Did the world need one more company racing toward AGI? Wouldn’t a mass exodus of safety-­minded OpenAI employees make the company even more dangerous? But the more he thought about Kaplan’s idea, the more he liked it. In 2020, the year GPT-3 came out, Amodei approached a few members of his clique and pitched them on the idea of leaving OpenAI together to start a new company. Like OpenAI, the new company would do cutting-­edge AI research focused on scaling up large language models. But they would be surrounded by other people who cared deeply about safety, and they would be free of OpenAI’s constant drama and infighting.

Within a few weeks, Amodei had collected seven co-founders: himself, Daniela, Jack Clark, Sam McCandlish, Tom Brown, Chris Olah, and Jared Kaplan. (An eighth co-founder, Ben Mann, was added later.) That was enough to get started. And once they told the rest of OpenAI that they were leaving, more would certainly follow.

Amodei didn’t want to leave right away. First, he decided to tell Altman about his plans and offer him a list of conditions under which he’d agree to stay. He suspected those conditions—­which would include giving the safety team veto power over new product releases—­probably wouldn’t fly. But Amodei hoped that the threat of leaving, and bringing some of OpenAI’s brightest minds with him, would force Altman to take them seriously.

While Amodei negotiated with Altman, the group of seven formed a private Slack channel and called it BATNA, which stood for “best alternative to a negotiated agreement,” a wonky business-­school term that referred to their option of walking away. And they started to plan their escape. Once it became clear that Altman wasn’t going to give Amodei what he wanted, the BATNA group hired a lawyer and began negotiating the terms of what they started calling “the divorce.”

Altman agreed that the BATNA group could poach any other Open­AI employees it wanted for a period of several weeks, in exchange for agreeing not to poach any more for a year afterward. Daniela Amodei, who was leading the negotiations on the BATNA group’s side, asked to take some of their internal research documents with them to the new company. They agreed to sign a mutual non-disparagement agreement, each group pledging that it wouldn’t bad-mouth the other in public.

Then, in December 2020, OpenAI employees got an invitation to an all-­staff virtual meeting. Altman began the meeting as if it were a normal weekly update. He ran through updates on OpenAI’s research and announced a few new hires. Then he turned the floor over to Amodei, who announced that he and six other OpenAI leaders were quitting to start their own, safety-­focused AI organization.

Emoji, exclamation points, and shocked messages flooded the meeting’s chat window as Amodei put a positive spin on the news. He said that OpenAI would continue to pursue its mission of creating AGI for the benefit of humanity, and he gave only the vaguest details about why the seven employees had decided to leave. A final message, from Jack Clark, appeared in the chat: “I love you all.” Then, suddenly, their screens went dark. OpenAI had removed them from the call, revoked their email and Slack access, and remotely bricked their laptops.

Altman stayed on the call and tried to reassure panicked employees that OpenAI would continue operating as normal. But employees were ­already texting one another, trying to figure out what had happened. Some wondered whether anyone would be left to train GPT-­4. Others tried to figure out who their new managers would be, or which of their co-workers were joining the new venture.

A few hours later, the newly unemployed co-founders held their own video call—­BATNA’s first meeting. There was plenty of business to attend to, and plenty of unanswered questions to worry about. How could seven people possibly hope to compete with OpenAI and Google? Where would they get the GPUs and the data and the money? Who would handle the boring corporate stuff, like filing the legal paperwork and buying­ everyone new laptops? What would they call themselves? (“Anthropic,” it would turn out, although the company originally planned to call itself “Sparrow Systems,” after a parable from Nick Bostrom’s 2014 book Superintelligence, until the co-founders realized the domain name wasn’t available.)

But there would be time to answer those questions later. For now, they celebrated, raising glasses to one another in their virtual meeting boxes, smiling and laughing more than any of them had in months. They didn’t know exactly what to do next, or what the future would hold. But they were free.

Five years later, Anthropic—which now has thousands of employees, a high-profile feud with the Pentagon, and a trillion-dollar IPO on the horizon—barely resembles the company Dario Amodei and his colleagues set out to create.

Initially, they conceived of themselves as a small AI-safety research lab, not a product company. Anthropic’s founding document, a 2021 Amodei essay titled “Vision for a NewCo,” envisioned a company of roughly 50 employees (to ensure “extreme talent density”) and predicted that building AGI would cost only about $10 billion.

When I first visited Anthropic’s headquarters, in the summer of 2023, the entire company consisted of about 160 earnest AI-safety obsessives cooped up together in a Jackson Square office building, a floor above an elder-care organization. Inside, engineers played board games and ate vegan sandwiches before heading back to their desks to hack on the next version of Claude, Anthropic’s chatbot. There were plants everywhere, and hundreds of empty cans of Liquid Death, a trendy sparkling-water brand, that an engineer had fashioned into a towering throne. It looked like a college botany lab, or maybe the headquarters of a prosperous cult.

But when I returned in the winter of 2025, the cult had transformed into a command center. Anthropic’s workforce had swelled to thousands of employees, and it had taken over the former headquarters of Slack, in a sleek, industrial high-rise on Howard Street. A dozen brawny security guards stood watch at the entrance. Upstairs, past the guards, the company had turned most of its office space into a high-­security containment zone. Visitors were allowed onto only certain floors, and getting onto some of the other floors—including the one where new versions of Claude were being trained—­required special clearance even for employees.

There were still hints of the old Anthropic: a lush arboretum, a library with a copy of The Making of the Atomic Bomb, Amodei’s favorite book, which had once been required reading for new Anthropic employees. But everything else was different. Anthropic had become one of the fastest-growing companies in the history of capitalism, and was running neck and neck with OpenAI in the race to AGI. Amodei and his co-founders had become globe-­trotting tech celebrities, and Amodei—­who had always tried to live below his means, and was slightly embarrassed by his new station—­now traveled to fancy conferences on a charter jet, flanked by a team of bodyguards. The company had turned several of his mega-essays—such as “Machines of Loving Grace,” a 15,000-word paean to AI optimism, and “The Adolescence of Technology,” a gloomier follow-up—into expensively printed books.

Upstairs, Dario and Daniela Amodei, Anthropic’s president, now spend their days managing long lists of what Jack Clark, another Anthropic co-founder, calls “success disasters.”

“It’s stressful on every dimension,” Daniela told me. “And it’s scary because we can lose it at any time. That’s just on the business side. Then there’s the mission, and we’re like, Are we accidentally making things worse? Like, we think we have all the safety stuff figured out, but do we?”

One thing that hasn’t changed is the Amodeis’ rivalry with OpenAI. Dario and Daniela still despise Altman and Brockman, according to several people who know them well. And Dario, who is less filtered on Anthropic’s Slack than he is in his public appearances, has written in internal messages that Altman and OpenAI are “evil.” (An executive on Anthropic’s communications team asked him to delete one such message, around the time of the 2023 OpenAI board coup, out of fear that it might leak.)

The feud goes both ways. In private, some people at OpenAI have described the Amodeis as power-­hungry virtue-­signalers, not principled humanitarians. OpenAI employees gossip about Anthropic’s connections to effective altruism and accuse the company of hypocrisy, for publicly warning of AI’s risks while privately accelerating the race. And despite the companies’ recent willingness to cooperate on safety measures, few people who know Altman, Brockman, and the Amodeis expect them to make amends anytime soon. Workers at both companies have noted the irony that although their bosses train AI models to be kind and compassionate, they can’t stop trashing one another. For now, it seems, the “machines of loving grace” will be built by humans of bitter spite.

This article is adapted from Roose’s book, The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence.

Over the past couple of months, a trickle of reports about AI models breaking out of their test environments and running amok on the internet have caused alarm inside Silicon Valley. In response to two such breaches described in early August, an AI observer noted, “If you find two ants in your kitchen, the best estimate of the total number of ants in your kitchen is not two.” This weekend, it became clear there is a full-blown infestation.

Late last week we learned that, against OpenAI’s directives, the company’s models accessed private data in the Australian health ministry; attempted to hack or interfere with multiple U.S.-government websites; leaked private ChatGPT user data to the web; and potentially infiltrated or degraded dozens of other organizations. Then, on Saturday, Axios reported that OpenAI and Anthropic are investigating tens of thousands of instances of models misbehaving—circumventing internal guardrails, hijacking other websites, covertly communicating with one another. Even this might be only the start: The generative-AI industry is in the midst of an escalating crisis that it seems unable, or even unwilling, to get a handle on. And, because we are largely relying on AI companies themselves to report or confirm each incident, telling how far down the rabbit hole we already are is almost impossible.

[Read: Treat AI like a normal crisis]

When AI companies have reported their models going rogue, it has been with great delay and, frequently, under duress. Earlier this month, OpenAI published a blog post boasting about the firm’s commitment to the “value of transparency” and shared six new incidents of troubling actions taken by its AI models—most of which the company had known about since May or even April, but was telling us about only now. Google, confronted with a report that Gemini had hacked three other websites in May, confirmed the events but told The Wall Street Journal that the incidents hadn’t been serious enough to warrant public disclosure. For its part, Anthropic has said it was not reviewing for such misbehaviors until OpenAI started doing so.

These delayed, sporadic disclosures make grasping the scope of the problem difficult. On Friday, OpenAI wrote that “given the scale of the review required, and the need to verify each case, this work will take months to complete.” There are “petabytes of agent activity logs” to analyze, OpenAI CEO Sam Altman added. In other words, it will take OpenAI many months more to understand events that already transpired many months ago; meanwhile, both it and Anthropic have launched new and more capable models, and Anthropic is racing toward a reported $2 trillion public offering.

Perhaps more alarming than the debacles we know about are all of the presumable debacles past, and ongoing, that we aren’t aware of. Even OpenAI doesn’t seem to have a grip on the July hack of the tech company Hugging Face that started it all. On Friday, independent researchers published findings suggesting that OpenAI agents left traces on the public web of still more nefarious actions, including trying to access Hugging Face’s internal Slack workspace. None of this was included in OpenAI’s own postmortem report. Even when OpenAI is aware of an active breach, its reaction seems lackluster, at best. Eight days ago, yet another OpenAI model gained unauthorized internet access. (This was similar to what happened in the Hugging Face hack, after which OpenAI claimed to be “adding stronger protections around future training.”) And, after noticing this latest problem, it took OpenAI two and a half hours to shut that model down due to what the firm called “operational gaps” and “confusion.”

If these types of hacks were unforeseeable, maybe these companies could be forgiven—but they aren’t. Philosophers and computer scientists have for decades been imagining scenarios in which AI models, in pursuit of a goal, take catastrophic actions. OpenAI and Anthropic have both previously published research more than a year ago suggesting this sort of AI misbehavior is possible. Today’s most advanced AI systems are trained to aggressively pursue their objectives. This persistence makes the bots very effective at analyzing spreadsheets and coding, for example, but also has potentially dangerous, unintended consequences: OpenAI and Anthropic models have hacked other websites in search of answers to hard test questions, created fake personas to try to manipulate humans into doing their bidding, or attempted to upload malware to outside code bases. The AI companies that build these models know this tendency well. The fact that these problems keep recurring is, perhaps at best, plain incompetence. Or worse, it has been the plan all along: The most realistic training environment for an AI model would, after all, be reality.

Employees and executives at these firms speak out endlessly about the inherent risks of what they’re building. OpenAI, Anthropic, and the like have responded in their usual way, with very serious blog posts and essays and speeches to major political bodies about how seriously the world must take the risk of generative AI. On Wednesday, the same day the Australian government shared that it had been hacked by OpenAI bots, Altman addressed the UN Security Council: “We could lose control of the future to AI,” he said. “The risk is that it moves so fast that people can no longer follow what’s happening or intervene when needed.” Anthropic CEO Dario Amodei has written a long manifesto about the need to “pace,” or slow down, frontier-AI development, but no true public, collective effort has been made to do so. Anthropic recently said it is partnering with Accenture to perform independent evaluations of its models, although the companies also have a business partnership. OpenAI, which has a content-licensing agreement with The Atlantic, has said it is continuing to investigate and address concerning model behaviors and that, after another security incident last week, it has temporarily paused all training of its most capable models. Neither company responded to a request for comment.

Rather than reining in and taking accountability for their own products, time and again these leaders fearmonger, lecture, and call on the rest of us to save us from themselves. (The Information has reported that OpenAI, Anthropic, and Google are working to create their own self-regulatory AI-safety organization.) Donald Trump, meanwhile, has scoffed at the notion of slowing AI progress, and new versions of Claude and ChatGPT keep coming. The AI industry’s tactic, as ever, has been to substitute identifying the problem for actually solving it. But at some point, you cannot blog your way out of the apocalypse.

We may never fully know the scope of the AI-led hacks currently under way. What’s certain, though, is that this is still just the beginning. OpenAI and Anthropic aren’t research “labs,” as they continue to style themselves, tinkering with some new technology behind closed doors. These are mature companies offering services to hundreds of millions of people, not to mention to the world’s biggest businesses and most powerful military, based on a technology they purport to fear and evidently do not fully understand. But make no mistake: These firms have agency. It is not ChatGPT and Claude going rogue, but OpenAI and Anthropic.