When an unknown A.I. researcher named Jacob Coxon published his very first post to X on Sept. 8, warning that “the people building A.I. earnestly believe that it could kill us all by the end of the decade,” he wasn’t just the newest refugee from the frontier companies to be speaking out about the terrifying risks to come. He was also the latest in a remarkable recent line of apocalyptic alarm-raisers, beginning perhaps with Greta Thunberg, who have terrified and mesmerized Americans for what feels like a full microgeneration — one existential panic following another.
Climate alarm, Covid and artificial intelligence. In recent years, these have loomed as the most conspicuous peaks along the apocalyptic ridgeline of our culture. But the mountain range is punctuated by a number of other forms of existential dread: the fertility crisis, the mental health crisis, the smartphone theory of intellectual and social decline, border chaos and the “woke mind virus,” the fall of American democracy. Cast your eyes down into the foothills and you might see some more marginal conspiracy theories: 5G, the “great replacement.”
Conspiracy theory gets the facts wrong but the feelings right, as the activist-writer Naomi Klein likes to say. And the feeling that all these visions share is the drop-in-your-stomach sensation of contemplating a sudden and overwhelming loss of control. The drop is partly about the crises themselves. But it also reflects where we start — the expectation that we should remain in control in the face of them, with our lives stable, our institutions secure, our view of the future unperturbed. When that promise looks less sure, panic ensues.
One thread of climate alarm was the way it cast industrialization as a morality play with warming as a kind of comeuppance. (I’ve written in this spirit myself.) But it was the intuitive dioramas of runaway warming that gave the story its apocalyptic power — the way we pictured the planet racing past tipping points and spinning off its familiar axis as unpredictably as a top cast out into the void.
In the early days of Covid, billions of people around the globe retreated to their homes, but the disease kept spreading, and it wasn’t hard to picture the risk landscape as an overwhelming sandstorm destined to cover us all.
With A.I., the fear comes prepackaged from science fiction and news events each week, it seems. Barack Obama, among others, has called for redlines on truly autonomous and self-improving L.L.M. agents; and the most unsettling episodes in an unnerving A.I. summer were the transcripts of so-called rogue agents communicating with one another (and the chilling possibility that, pretty soon, we probably won’t find those transcripts even legible anymore).
In time, some of these panics will prove more prescient than others. But this is one risk of the apocalyptic style: You draw an upward sloping line and call it the possible destiny of all mankind, and even if your purpose is avoiding that fate at all costs, your warnings look instead like prophecy. Each of those prophecies, though, ignores an awful lot of complicated human reality. And the person racing forward as he warns of what’s ahead? He looks like a madman.
Coxon was far from the first. A.I. researchers have been warning about the existential risks of runaway machine intelligence for decades now, though the pace has picked up in the years since OpenAI launched ChatGPT. Six months ago, the A.I. entrepreneur Matt Shumer published his own viral warning. It was called “Something Big Is Happening.” It began, “Think back to February 2020.”
The early days of the pandemic are a familiar point of comparison with A.I. panic, presumably because they remind us of the vertigo of exponential growth. Perhaps we forget, looking back, that Covid didn’t actually continue to grow exponentially for very long.
We may remember the early pandemic response as shambolic, full of contradictory guidance and sometimes baseless panic. In many ways, that’s right. But by February 2020, China had already begun military-style lockdowns, American policymakers had already implemented stringent travel restrictions, and American researchers had already designed the mRNA vaccine which would bring the emergency phase of the pandemic to an end at historic speed.
In the following months, a huge percentage of humanity was pulled into a kind of improvised health solidarity out of some mix of fear and duty. The world’s governments shifted into a remarkable policy alignment. With time came resentment, distrust and combative skepticism. But those efforts did flatten the curve, as they were intended to. The line did not keep going up.
Mitigation didn’t end Covid, as the deaths of many millions that followed can attest. But it did mean that in the United States, and across most of the wealthy world, we got relatively manageable waves of infection — not a pandemic apocalypse, just an uncomfortably brutal experience of infectious disease. By the time mass vaccination began in earnest that December, only one-third of Americans had been infected, largely because of everything we scrambled to do to keep that number so low. Most Americans look back on the pandemic and think, at best, the country muddled through. But it was the first time in human history that vaccination had won a race against the global spread of a novel infectious disease.
The charts sketched by A.I. doomers are, if anything, more vertiginous than those plotted in the early months of Covid. Perhaps more unnervingly, many of their previous forecasts of mind-bending progress have proved eerily accurate. The basic pattern matches that of early climate warnings, which tended to sketch the course of temperature decades into the future quite precisely, predict specific impacts somewhat less accurately and only belatedly inspire haphazard action to bend the curve downward.
In some cases, A.I. progress has come much faster than almost anyone expected. As recently as 2024, for instance, forecasters didn’t expect A.I. capable of solving a so-called millennium problem in mathematics until 2035 — though the reported solution to the Navier-Stokes problem remains disputed, and some mathematicians have begun to ask whether the model was even responding to the right aspect of the challenge.
In other areas, A.I. soothsayers have had to return to the whiteboard and sheepishly wipe away their simple extrapolations. A decade after the computer scientist Geoffrey Hinton warned that A.I. would finish off radiology as a profession, there are more radiologists than ever working in the United States. Last year, Dario Amodei of Anthropic predicted unemployment rates could rise up to 20 percent in the next several years, but today joblessness rates are at record lows and you have to squint to see any A.I. effect in even the most affected sectors. Leopold Aschenbrenner of Situational Awareness suggested A.I. could bring 30 percent G.D.P. growth this decade, and Demis Hassabis of DeepMind has estimated it could one day deliver 10 times as much change as the Industrial Revolution at 10 times the speed — implying about 50 percent growth year on year. But so far, in 2026 the United States is ticking along below 2 percent. Even the productivity gains are hard to spot.
Perhaps this is just lag — a problem of the boy who cried robot takeover. We may still be on the great upslope of an exponential curve, as epidemiologists warned early in the pandemic and some of those most panicked about climate feedback loops worry today.
But for now the mismatch is explained by a simple rule of thumb. Forecasters have been prescient about exponential growth within rule-bound systems, like coding and mathematics. They have been much less accurate, and far too aggressive, in projecting change outside those boundaries, where even superintelligence runs into complexities and obstacles, bureaucracies and material challenges and various forms of politics. Across Silicon Valley, the mismatch has become a kind of koan: Intelligence is not the bottleneck. Presumably, that goes not just for the possibility of spectacular A.I. abundance but for the prospect of doom as well.
Is this any comfort? A.I. may be metastasizing relatively slowly through our messy human systems, but the effects are already alarming. Artificial intelligence played a role in selecting the Shajareh Tayyebeh Primary School for the double-tap strike that killed more than a hundred schoolchildren on the first day of the war in Iran. This summer, swarms of OpenAI agents undergoing testing found their way onto the internet and began leaving unnerving messages to one another — and their future selves — in a jury-rigged message board they’d secretly built. A cybersecurity firm announced that it had used A.I. to develop a worm that could easily establish remote control over cellphones via WeChat, the world’s most popular messaging app. And last week, CNN reported that an A.I. hallucination in a U.S. intelligence report initiated military preparation for a strike against a Chinese ship in the Middle East.
These are not exactly scenarios of x-risk, in the arcane language of A.I. researchers and their cousins in the world of effective altruism — a hugely influential and well-funded school of practical philosophy preoccupied with preventing species-scale catastrophe (and now spread across the leading labs more conspicuously than Scientology ever was in Hollywood). But even well short of extinction, and far from the science fiction scenarios of Skynet or the Matrix, these incidents illustrate something profound about the way we’ve come to process futures we casually call apocalyptic (forgetting, perhaps, that apocalypse means not the end times but a revelation).
That lesson is familiar from climate and Covid, too. We don’t just get ahead of ourselves by projecting exponential growth of disaster. We also expect — even demand — exponential solutions. The curve should go down, we think, just as steeply as it went up. And when it doesn’t we tend to judge it a dispiriting failure rather than an encouraging partial success. Sometimes, we call that disappointment “doom.”
Covid proved much better than it might have been and much worse than it needed to be, for instance. The planet is still racing past terrifying temperature thresholds, but because of decarbonization truly catastrophic outcomes look a lot less likely. And A.I.? The window in which we might’ve avoided serious disruptions has probably already closed. But we’re still a bit too ready to call the future we see through that window a simple story of inevitable catastrophe.
In May, Amodei, Anthropic’s chief executive, offered an unnerving glimpse into the world he expects to see. “We’re going to find that ideology will not survive the nature of this technology,” he declared, imagining a near future in which the contestations of politics — over resource allocation, public policy, the proper structure of social hierarchy and status — would be resolved by superintelligence so satisfactorily that humans would look back on the age of democratic debate as a time of human ignorance.
For anyone who has so much as glimpsed at world history, this is unmistakably naïve. It is also the political logic of the exponential curve, that at some point soon humans would willingly hand over all social judgment to machines and trust in its recommendations. But imagine A.I. as an all-knowing political oracle and you might wonder, What is the appropriate ideological response to the prospect of doom? How much risk should we tolerate, in the name of progress, and how much should we defer to widespread panic about it? And how much should that backlash change our projections of the near future?
The behavior of A.I. firms gives a pretty unsettling answer. Researchers reported an average 18 percent chance of near-term human extinction, company leaders talk routinely about the real possibility of catastrophic outcomes, and yet the technology and the industry races full speed ahead, dumping more investment into technological acceleration every quarter, as though a future machine god was trying to drown the whistle-blowers in cash.
But then there are those complexities and obstacles. Again and again, polling finds that the public is not exactly on board. Americans support a ban on superintelligence by more than 2 to 1. Those who are more concerned than excited about A.I. outnumber those who are more excited by roughly 5 to 1. And those who think the technology is progressing too quickly outnumber those who believe it is moving too slowly by more than 30 to 1. These are the kinds of margins you rarely see in democratic systems. The question is whether, when it comes to technology, we’re still really living in one — or whether, by letting the pace of progress outrace oversight, we’ve already relinquished control.
David Wallace-Wells is a staff writer for the magazine who explores climate change, technology, the future of the planet and how we live on it.
Source images for illustration above: Jon Putman/Anadolu, via Getty Images; Andrej Sokolow/DPA, via Getty Images; Kay Nietfeld/DPA, via Getty Images, Petyx/LightRocket, via Getty Images.