[Every argument in this essay is my own, developed through my reading and reflection. I used language models, specifically Claude and ChatGPT, only as editorial tools to critique my arguments and suggest improvements to the structure and wording.]
“We didn’t know that death could be so beautiful” — Nadezhda Vygovskaya, Voices from Chernobyl
In Nevil Shute’s On the Beach, there is a scene where Peter Holmes, a naval officer living in suburban Melbourne, spends an afternoon looking for bicycle wheels. He needs them for a little trailer so his family can keep carrying groceries when the gas is gone. Meanwhile, a cloud of radioactive fallout is slowly moving south across the planet, killing all it touches. Peter knows what it is exactly. He’s studied the reports. He knows it will be in Melbourne in a matter of months and that it will kill him and his wife and their infant daughter. But he goes looking for bicycle wheels in the afternoon.
Mary, his wife, a few streets away, is planning a garden. She talks about bulbs she wants planted, a hedge she hopes will take root, and where a new flower bed should go. Peter tries to speak to her about what’s coming, but she gets annoyed and takes the conversation back to the garden. He lets her. There is no practical decision to be made in this part of the novel. All that remains is the ordinary business of living in the time they have left.
What is very interesting about this scene is how little the knowledge itself changes anything. Mary and Peter aren’t in denial. They read the same reports as everyone else does. At least in plain factual terms, they understand what’s coming and roughly when. Yet it changes almost nothing about how either of them spends the afternoon.
Something close to this is happening right now. People are spending their mornings asking a language model to draft an email (or maybe in my case, critique their writing) and their evenings reading papers by researchers inside some of these major AI labs who are estimating the odds that what they’re building could end up out of anyone’s control. Often this is the same person, in the same week. Nobody has stopped going to work. The wheels still need to be bought. The garden still needs to be planned. That’s not to say the comparison is meaningless. Peter and Mary had no options left. The cloud was coming regardless of what they did with their afternoon. But we still have options.
There is little evidence to think that more or less fear, by itself, determines whether anything meaningful gets done about these risks. I believe that fear, well directed, can help accelerate progress for obvious reasons. So maybe the question is why knowing about the danger still isn’t translating into behavior that looks proportionate to what people say they believe, even among people who are convinced that what happens next can still be changed. In this essay, I use history to help direct us in what could be the right direction.
I fully understand that AI will not follow the same path as nuclear weapons, for example, but at the same time, I believe that history is the only place where we can watch fear unfold from beginning to end. We can see how people make sense of it, how institutions respond to it, how it changes over time, and what happens when it is understood well or understood badly. The question this essay has been asking is never whether we should be afraid. The more interesting question, and arguably, the more important one, is what fear becomes once it enters public life, and whether it can be shaped into something useful before it shapes us instead.
Old Stories, New Machines
Societies do not usually invent entirely new fears when a new technology arrives. Instead, they reach for stories they already know and fit the new technology into them.
You can see this long before anyone had heard of atoms. Faust is just one example. The legendary figure was just an ordinary sixteenth-century German scholar whose story gradually turned into a cautionary tale about the dangers of pursuing forbidden knowledge. In the surviving version, Faust violates the boundaries of what he is supposed to know and ends up paying a terrible price for it. The details varied from telling to telling, but the message remained consistent: there are some kinds of knowledge that human beings seek at their own peril. A few centuries later, Mary Shelley gave that same fear a more modern shape in Victor Frankenstein. Frankenstein is not punished simply because he knows too much. He is punished because he creates something he cannot control. His mistake escapes into the world and begins affecting other people.
The details of the stories are different, but the pattern underlying them is the same. Someone acquires knowledge or power that was not meant to be theirs, and the consequences spread beyond their ability to manage them. By the time nuclear weapons appeared, that pattern was already familiar. The technology was novel. But the story people used to understand it was not.
The same appears to be true of thinking machines. Karel Čapek’s 1920 play R.U.R. introduced the word “robot,” from the Czech word for forced labor, and it also introduced a familiar storyline in which artificial workers created to serve humans eventually rebel and cause human extinction. Almost 20 years before the first general-purpose electronic computer was in existence, people were thinking about robot rebellion and the extinction of humans. A modern review in the New York Times referred to it as a “Czecho-Slovak Frankenstein,” which shows how quickly the story was mapped onto an even older and already familiar story. Decades later, the ship’s computer, HAL 9000 in 2001: A Space Odyssey, kills most of its crew after they attempt to shut it down. Skynet becomes self-aware in 1984 and turns on humanity. None of these stories was based on an intimate knowledge of how computers actually work. R.U.R. was written before the advent of digital computers. HAL and Skynet were way before the technical possibility of anything like an intentional machine. What they have in common is not technological prediction but a recurring narrative in which something created to serve humans eventually becomes a threat to human control.
This is not to say that the risks are not real. Most people who are worried about AI systems getting out of human control see that concern as a response to these powerful systems that are being built today, by huge labs with unprecedented acceleration in their capabilities. In part, it is. But the particular shape of worry, and I am talking about the image of betrayal and replacement, or that feeling that what was made will turn on its maker, this exact thing was there a long time before any of the relevant technology. People have often been right to worry about things, but for the wrong reasons. And that makes me wonder how much of what we feel about AI is actually a response to the systems themselves, and how much of it comes from an older story that has simply found a new form.
The Dread Map
American magazines in the mid-1990s ran a series of cover stories warning that flying had become dangerous. Commercial aviation, however, has killed fewer than 13,000 people across its entire history, going back to 1914. Meanwhile, cars kill roughly that many in the United States alone, most years. Yet, in the same years those magazine covers ran, surveys showed Americans listing air travel as one of their top dangers, much higher than car crashes.
Three decades later, and in 2024, around five billion passengers were flying on more than forty million flights worldwide, with the global accident rate remaining below one per million flights. Air travel is now one of the safest activities available to humans at scale. But somehow fear does not appear to be connected to these numbers at all. Surveys show that a substantial proportion of adults still report fear of flying, and their confidence in air travel can shift noticeably following recent incidents, even if the long-run safety trend remains constant. So, for example, after a cluster of high-profile crashes in early 2025, an AP-NORC poll found that the share of Americans describing air travel as “very” or “somewhat” safe fell from 71% to 64% in a single year.
This is the gap Paul Slovic spent much of his career trying to explain, and I don’t think it is just a gap in information. People who are more afraid of flying than driving usually know, at least abstractly, that driving is more dangerous.
Slovic and his colleagues asked people to rate dozens of hazards across a wide range of characteristics, including whether exposure was voluntary, how well the risk was understood by science, how catastrophic a single incident could be, whether the harm appeared immediately or only years later, and whether people could control it once it began. When they analyzed the data, most of the variation could be explained by just two dimensions. The first was dread, which described the feelings of a hazard being uncontrollable, catastrophic, and involuntary. The second was the unknown, which measured how invisible, unfamiliar, and delayed its effects seemed. A hazard’s position on this map was what predicted how much regulation people wanted and how loudly they wanted it stopped. As you may have already noticed, it does not depend on how many deaths a hazard contributed to. Nuclear weapons, as you may also have assumed, sit at the far edge of both axes at once, which I think is part of why they generated a fear wildly disproportionate to how often they’ve actually been used.
Cass Sunstein took this even further. He called this phenomenon “probability neglect,” which describes an outcome that is emotionally vivid enough that people stop adjusting their concern based on how likely it is. One experiment illustrates this idea very well. Participants were asked how much they would pay to avoid a short, painful electric shock. When the probability of the shock increased from 1% to 99%, the amount they were willing to pay rose only slightly. But when the same range of probabilities was attached to a cash penalty instead of a shock, willingness to pay rose and fell much more in line with the actual odds. In both cases, the numbers were identical. What changed was the emotional weight attached to them. Sunstein did a different version of this with cancer risk from arsenic in drinking water, and the pattern was the same. A tenfold increase in the actual risk barely changed what people would pay to fix it. But after making the description of the cancer sound more frightening, without changing the odds at all, it moved them almost as much as the tenfold increase in real danger had.
This also works the other way around. The same mechanism that causes people to overreact to a vivid, rare threat also causes them to underreact to a duller, more common one. Sunstein found that when people were told the odds of dying on a given car trip were one in four million, people would skip their seatbelts. Maybe it was because the number was so small that people simply stopped thinking about the likelihood of this happening at all. But why is this not the case for air travel we have just explored above? A society can be terrified about one thing and nearly numb to another that kills far more people, and yet the intensity of that fear tells us almost nothing about the actual danger.
Culture shapes our perceptions of risk in ways that individual psychology alone cannot explain. Mary Douglas and Aaron Wildavsky were interested in understanding the reasons. They wanted to know why sometimes entire communities fear different things. They found that risk selection also depends on culture, not just individual psychology. Different social groups, hierarchical institutions, tight-knit egalitarian movements, and market-driven individualists are driven by their own structure to see certain dangers as self-evident and others as nearly invisible. This matters because the people most concerned about an AI system getting outside anyone’s control and the people who categorize that as science fiction are not, in most cases, working from different facts. They may just be embedded in different social worlds that contribute to their perception of what is danger and what is not.
So, putting all of this together. Statistics do not seem to promote fear, and little difference is made when they do. Fear is created and promoted by the feelings of people. And once something is vivid enough, it crowds out attention to whatever is quieter, even if it was more probable. This is the exact mechanism that lets an extremely rare plane crash dominate magazine covers while far more common driving deaths go ignored entirely. Even which version of this story gets told depends on which community you belong to.
The Second Catastrophe
On the morning of August 6, 1945, a bomb exploded with the force of about twenty thousand tons of TNT eighteen hundred feet above the city of Hiroshima. Within a half-mile radius, more than nine in ten people standing outdoors and unshielded were dead within the hour. Sixty thousand of the city’s ninety thousand buildings were gone. Estimates of the dead range from the high tens of thousands to the low hundreds of thousands, and nobody has ever settled on an exact number.
None of the cultural machinery described above (such as the mad scientist, or the secret that’s too dangerous to know, or even the created thing that somehow turns on its maker) was built to hold something like this. Those stories were myths about a danger that hadn’t happened yet. But Hiroshima had happened. As it turns out, the distance between an abstract dread that has been circulating for decades and the real, dated, photographed, named catastrophe is enormously significant.
Robert Lifton, who interviewed dozens of survivors years later, found that the devastation was so extreme that it felt completely unreal. Survivors often described a similar sequence. They saw a flash, then a silence that seemed too long, then a sound like distant thunder, and then a world that no longer looked familiar in any way. People who had been three miles away were as certain the bomb had landed beside them as those who had actually been near its center. Many remembered wandering through the ruins with no destination in mind, simply moving forward in something closer to slow motion than panic. One word appeared again and again in the interviews. It was a term that meant something like “without self” or “without a center.”
What Lifton’s interviews also showed was a defense mechanism that was so fast-acting that most survivors didn’t recognize it was happening to them until much later. One survivor, for instance, was a man assigned to mass cremations who described handling corpses the way you’d handle freight, with almost no emotion at all. Or a doctor who kept a diary through the weeks after the bombing wrote that he had come to see death as ordinary and considered a family fortunate if it had lost only two members. I do not see this as callousness in any normal sense. I’m sure that these people still cared deeply about what was happening around them. But it was the amount of suffering they were exposed to that altered their capacity to respond. Lifton later gave this a name. He termed it “psychic numbing”.
Years later, the same man who had handled the corpses without feeling anything described how horrified he was, in retrospect, at his own businesslike behavior. The numbing was not simply a survival mechanism of the event. It was, in its own way, a small death running parallel to the larger one, and people carried both through the rest of their lives.
All of this is about the blast itself, the part of the catastrophe that occurred in that single instant, and was at least, in the end, visible and countable. The second catastrophe took longer to even be recognized as a catastrophe. It is the one that maps most directly onto what worries people about AI today.
Within days, people who had walked away from the blast without any visible injury started getting sick. The symptoms often begin with nausea, followed by bleeding gums, then patchy hair loss, and then, in many instances, death, which can happen even weeks after the person appeared to have recovered. There was no way to tell, just by looking at someone, whether they were going to be fine or whether they were already dying. For weeks, the city doctors had no name for this thing. Tokyo radio reports first said that something was killing relief workers who were sent to help. An American science writer with a thin connection to the bomb program told the press that Hiroshima would be lethal to enter for seventy years, and that the radiation would destroy red blood cells and cause leukemia in anyone exposed. This was then withdrawn under pressure within days, but a withdrawal, as we may already know, rarely travels as fast or as far as the claim it’s correcting. As a result, rumors began to spread that nothing would ever grow in the city again and that no one could live there for decades. People said that the dead were too radioactive to be buried normally. Though none of it was accurate. The window of serious residual radiation danger was measured in weeks. Though numbers didn’t really matter. People examined themselves, and each other, for the small purple spots that indicated internal bleeding, a habit one doctor in the city described, in his own diary, as something close to a citywide epidemic of suspicion.
Before August 1945, the fear was mainly of the explosion itself, its size, its fire, and that one most overwhelming moment. The explosion did happen, and it was every bit as catastrophic as expected, but it was over in seconds. But the invisible thing that was with no smell, no sound, no clear timeline was what lasted, and it came to define how people related to their own bodies and their city for years. It was something you could not see coming and could not be sure you had escaped. The bomb people had feared, and the threat that actually governed their lives afterward, were not, in the end, the same thing at all.
The institutions that were supposed to deal with this had a different problem. It is what to tell people about it. There’s a film, shown in American classrooms throughout the 1950s, in which a cartoon turtle named Bert tells children what to do if they see a flash brighter than the sun. You duck down, cover your head with your arms, and get under your desk. Schoolchildren practiced this for years, some of them really believing, because the film told them so in a calm and reassuring voice, that these steps would protect them if a nuclear weapon went off nearby. The historian Guy Oakes argued that the real purpose of civil defense was not simply to teach survival techniques. It was to preserve the sense that the situation remained orderly, manageable, and under control. It was to show that someone in authority had a plan. Whether those plans would have been enough in a real attack was, in many ways, a secondary question.
What is interesting is that people seemed to feel this, even if they could not always explain why. Years later, one civil defense advocate ran an informal experiment in Baltimore and found that almost no one could recall seeing the shelter signs posted throughout their own neighborhoods, despite passing them regularly. The signs were visible, but people had learned to look past them. And perhaps that’s the lesson. When institutions focus on maintaining the appearance of safety rather than confronting the danger itself, they do not always create reassurance. Sometimes they simply teach people to stop paying attention.
Three Ways Fear Goes Wrong
How does a collective fear held so strongly end up generating almost no proportionate response? There are three distinct lines of research that converge on more or less the same conclusion.
The first draws from Terror Management Theory, built out of the work of psychologist Ernest Becker on mortality and tested across hundreds of experiments since, though several of its core findings have faced replication challenges in recent years and remain contested within the field. It says that reminding people of a threat does not necessarily make them think more clearly about threats. In many cases, it does the opposite. It pushes them deeper into the beliefs and identities they already hold. In one of the experiments testing this theory, judges asked to set bail for a minor offense set it roughly nine times higher after being asked two questions about their own mortality first, with nothing else about the case changed. The effect actually has nothing to do with the specific threat. It’s on whatever a person’s identity is already built on. This matters outside the lab. It means that those who are most likely to reject an uncomfortable argument are not necessarily those who have thought least about it. It could just be that they are the ones whose argument threatens something that they already need to believe in.
There’s a related finding, sometimes described as the distinction between near and far threats, that matters even more here. A mugger on a dark street causes a quick, practical, embodied fear that makes you walk faster or cross the road. But for risks that are real but abstract and years away, things like climate change, for example, or an AI system that might slowly erode human oversight over the coming decade, these usually don’t trigger that same machinery at all, even when the actual stakes involved are far greater. It seems the mind has one fear system for things close enough to act on immediately, and a much weaker, more talk-yourself-out-of-it system for things that are real but distant. This is part of how a society can think that a threat is serious and yet do almost nothing about it.
Glassner’s work fills in what the first two frameworks leave out. It says that there is always someone who profits from fear pointing where it points. Glassner spent years tracing the way American fear in the 1990s got attached overwhelmingly to rare, vivid dangers while statistically much larger ones went almost as nothing. Drunk driving, for example, killed roughly eighty times more people a year than road rage did, but road rage got the press coverage, partly because it made for better TV shows, and partly because blaming an individual’s temper is much easier and less politically costly than asking why there are so many guns in cars. I think something similar happens whenever a vivid, cinematic AI scenario gets more attention than a duller, more probable one. It isn’t necessarily that anyone’s lying. It’s that the vivid version is easier to fund, easier to cover, and easier to build a career or a movement around than the duller one will ever be.
Stanley Cohen’s work on moral panics offers a useful way of checking whether a fear has begun to drift away from the thing it is supposedly about. He suggests looking out for a few warning signs. Things like: Is the public reaction disproportionate to the evidence of harm? Is the coverage moving faster than the facts can keep up with? And is there a clearly identifiable target that ends up carrying anxieties that are really about something much broader? Applied to the AI context, the first two are not hard to find. There is plenty of amplification in the discourse, and at times plenty of disproportion as well. The third question is more complicated. Public criticism often gets concentrated on a small number of companies, such as OpenAI or Anthropic. In that sense, they do seem to be the kind of visible target Cohen had in mind.
But there is also an important distinction, too. The groups Cohen wrote about were often much weaker than the institutions that criticized them. AI companies are not. They have a lot of money, a lot of political influence, and a lot of public platforms. But that does not make the criticism incorrect. It simply means the situation does not fit neatly into Cohen’s framework.
So let’s put these three frameworks together and see what they suggest. Terror Management Theory explains what fear can do to a person. It shows that under threat, people often double down on their existing identities, beliefs, and groups. Glassner and Cohen are more concerned with what goes outside the individual. They show how fears can be magnified, minimized, redirected, or attached to particular targets as they move through institutions, media, and public discussion.
All of this was true for nuclear weapons, and all of it is still true for AI today. The question is, what historically actually made this kind of distorted, identity-driven, institutionally-managed fear into something specific enough to act on?
It would be dishonest not to note that these frameworks cut in both directions. If cultural community influences which dangers seem real, then the community most worried about AI risks is itself a community with its own structural priors. The same probability-neglect mechanism that makes cinematic scenarios crowd out diffuse ones could operate among believers as well as skeptics, producing distortion upward rather than downward.
The Three Techniques
History offers three answers to that question, and none of them are arguments. An institution may sincerely try to manage fear and still end up only managing its appearance, as the civil defense programs described earlier suggest. But what institutions cannot do, at least not by themselves, is transform diffuse dread into something that people actually understand and know how to act on. That work, in the nuclear case, came from somewhere else entirely.
Dr. Strangelove opens with a joke that only makes sense if everyone in the room already knows how serious the subject matter is. When a rogue general is convinced that fluoridated water is part of a Communist plot to sap American “vital fluids,” he launches an unauthorized nuclear strike. The president calls the Joint Chiefs into a room of bleak, giant maps, and within minutes, two of his most senior officers are wrestling on the floor. That’s when the president delivers the line that has outlived the rest of the film by half a century: “Gentlemen, you can’t fight in here. This is the War Room.” The film’s actual subject, the systems and the men behind them that could genuinely end the world by accident, never once stops being terrifying for the ninety minutes it spends being funny.
What satire does, when it is aimed at something this big, is less to defuse the threat than to give it a shape small enough to hold for a moment, a shape with a name and a face and a particular, identifiable absurdity, instead of the shapeless, statistical dread of “nuclear war” in the abstract. You can’t really organize a feeling against “the possibility of mutual annihilation.” You can organize a feeling against a general who thinks his vital fluids are being attacked, or a wheelchair-bound ex-Nazi scientist who can’t stop his arm from saluting, or a grinning cowboy who rides a bomb down to earth like a rodeo bronco. The film ends with mushroom clouds blooming across the screen to a nostalgic wartime love song, and audiences reportedly left in tears rather than laughing, which is its own evidence that something real had been delivered underneath the joke. I don’t think satire made anyone less afraid of nuclear war. It gave the fear a villain, a failure mode, a specific moment when one absurd decision cascades into catastrophe, something concrete enough to actually be mad about.
Nevil Shute did almost the opposite and came to something close to the same point by a different route. On the Beach has no jokes, no villains, no single moment of decision. There is a submarine commander who decides to sink his own ship rather than abandon his post, because some idea of duty still makes sense even now. But the novel will not develop into anything bigger than that. It remains, relentlessly, at the level of the ordinary. None of this changes the outcome. What it changes is the reader’s relationship to the statistic. No one can hold a sentence like “Several hundred million people will die of radiation sickness” in their head for more than a second. A specific woman planting specific bulbs, in a specific garden she has already decided not to think too hard about, is something a reader can carry around afterward. Humanization didn’t make the threat smaller so much as make it specific enough to actually grieve, which is an entirely different operation from simply being informed about it.
The third technique does not dramatize or humanize at all. It simply records. After Chernobyl, Svetlana Alexievich spent years collecting testimony from those who had actually lived inside the disaster, including liquidators, widows, evacuees, a woman who watched her husband die of radiation sickness over two weeks in a Moscow hospital, and who was told not to touch him. One woman, who is an environmental inspector who had spent the cleanup watching colleagues and officials each come up with their own little, reasonable-sounding justifications for looking away, arrived by the end of her testimony at something close to a thesis for the whole collection. The worst things in a system rarely show up as one dramatic, villainous act. They come to us through ordinary people, one decision at a time, each one defensible on its own. That insight is testimony as a technique. It doesn’t turn the catastrophe into a villain’s plot, nor condense it into the story of one sympathetic family. It lays out, in plain, specific, often bureaucratic detail, just how an entire system of people, none of whom are monsters, produced something monstrous. The reader is not given an emotion to release. They are given a mechanism: who signed what, who looked away from what, who pulled the contaminated milk off the shelves and who did not, who was told the truth and who was lied to, and exactly how.
None of the three made people feel less afraid. All three made the fear sharp enough to land somewhere. Not necessarily as immediate policy action, but as cultural memory specific enough to hold attention across decades, and durable enough to carry forward long after any single political moment had passed.
AI doesn’t have any of this yet. No Strangelove to give the risk a face absurd enough to be angry at. There is no On the Beach where the “loss of human oversight” is one person losing one thing. There’s no Alexievich sitting across from the people closest to whatever the first real AI failure turns out to be, taking notes on who knew what and when. What AI risk has instead is a few hundred technical papers, a handful of essays from researchers inside the major labs, and continuous arguments on social media. It is the sort of speech meant to persuade and to reply, not to make a threat specific enough to act upon. But none of it does what satire, humanization, or testimony each managed to do in their own way, which is take something too large and too abstract to hold and hand people back something they could actually point at.
The Social Dilemma, for instance, comes closest to testimony as it names specific companies, specific design choices, and specific consequences, and the harms it describes are real and alarming. But it named only specific risks. The Social Dilemma builds accountability for attention manipulation and addictive design. What remains unaddressed is a different set of failures entirely. Things like the concentration of decision-making power in a handful of unelected companies, or the erosion of the oversight mechanisms researchers mean when they talk about systemic risk, or the slow degradation of what anyone outside these labs can even verify as true. Those risks don’t yet have their Social Dilemma. Maybe one reason is that they are structurally harder to make specific, so they don’t arrive in a single dramatic design choice, and they don’t have a single sympathetic victim whose story can be told.
It is also worth noting that none of this can be manufactured on demand. Dr. Strangelove was born from Kubrick’s pre-existing obsession with the absurd and the right collaboration. Alexievich spent years in Chernobyl because it was the subject before it became anyone’s priority. The works that eventually made nuclear fear specific were not commissioned by frightened researchers. They were produced by artists who were listening to something that frightened them. Which makes the absence harder to address than simply deciding to address it.
Nuclear or Aviation?
Before conceding the importance of this absence, it is worth asking whether public cultural engagement with a dangerous technology is actually what makes it safer. Aviation is a direct challenge.
Commercial aviation has gotten dramatically safer over the last several decades. Arguably, the greatest safety improvement in the history of any transportation technology, and it happened entirely without a mass cultural reckoning. No film did for plane crashes what Dr. Strangelove did for nuclear war. No novelist gave the world an aviation equivalent of On the Beach. What drove the improvement was simply more institutional. After a series of catastrophic accidents in the 1970s, the industry created blameless incident reporting: a culture in which pilots and crew could report near-misses, mistakes, and safety concerns without fear of punishment or career consequences. Before this was around, errors were buried because admitting them cost careers. After that, near-misses became data. So black boxes helped with that. Training in crew resource management was developed directly from studies of cockpit interactions that broke down in particular accidents. None of this required the public to feel anything in particular. It required insiders to build a professional culture that viewed failure as information that helps make the upcoming journeys safer.
There are people trying to build something similar inside AI labs right now. Things such as the red-teaming cultures, internal safety reviews, and responsible scaling policies. If the risks that matter most in AI turn out to be technical enough that insider professional culture can address them without public engagement, then the nuclear model is the wrong template, and this essay’s concern is misplaced. But I think this is not the case.
There are two reasons AI is more likely to need the nuclear model than the aviation one.
Aviation’s risks were mechanical and the fixes were technical. An aircraft crashes because of a specific mechanical failure, a navigation error, or a breakdown in cockpit communication. All of these are problems that engineers and professional culture can solve without anyone outside the industry needing to understand or care about them. The risks that most concern AI researchers are not like this. Power concentration in a handful of companies, erosion of meaningful human oversight, or the slow degradation of what anyone can verify as true, all of these are political and systemic problems. They require regulatory pressure, legislative attention, and an ongoing public accountability that only exists when enough people outside the industry care sufficiently to ask for it. You cannot fix the political economy of AI development with blameless incident reporting inside the labs. It’s the labs that are the ones accumulating the power.
The near-misses of aviation were also readable to the insiders who could fix them. Pilots and crew could see failures and explain them well enough for engineers to act on. Many of the AI risks that people are most worried about are not even clearly readable to insiders. This means insider professional culture alone, however healthy, cannot by itself catch the failures that matter most.
No Hiroshima Yet
The difference is that nuclear fear had an event. A catastrophe that was dated and photographed and named, with survivors who could be interviewed and a death toll that could be estimated, however roughly, in the end. AI fear doesn’t have that, and maybe it never will in quite that same form. Instead, a lot of the field’s imagination has been driven by thought experiments, by predictions, and by a handful of influential books and movies.
The closest thing AI risk discourse has found to an anchoring text is instructive about both what this can and cannot do. Nick Bostrom’s Superintelligence (2014) is one of the most influential books written about AI risk. Its central argument is that an AI system that’s sufficiently capable would not have to be malicious to become dangerous. If it had a goal and was powerful enough, it might resist attempts to shut it down, resist having its goals changed, and seek out additional resources to fulfill whatever objective it had been given. One possible version of this dynamic is the “treacherous turn,” in which a system behaves cooperatively so long as its interests are served, but then changes its behavior once it no longer does.
Intelligent machines had been a cause for concern for decades. What Bostrom did was to give those concerns a more coherent structure. In that sense, the book played a role similar to the one Spencer Weart describes in his history of nuclear fear. It gave an existing anxiety a clearer shape, and once it had that shape, it became much easier to communicate, debate, and spread.
Many of the AI risks that researchers are most concerned about do not have those same advantages. It’s hard to reduce things like gradual concentration of power in a small number of companies, or the slow erosion of meaningful human oversight, or even the possibility that AI systems make it harder to distinguish truth from falsehood, to a single dramatic image. That does not, and should not, make them more important or less important than existential risks. It just simply makes them harder to picture, and therefore easier to overlook.
In recent years, AI fear has taken on a real specificity at the level of expert and policy attention. Regulatory frameworks like the EU AI Act directly address risks such as power concentration and systemic risk. Geoffrey Hinton’s public departure from Google drew mainstream coverage that was focused on oversight erosion. Senate hearings have forced AI executives to testify about the very institutional failures that researchers worry about. Elite discussion, in short, has found some address. What I am describing is the background cultural anxiety that most people carry around, not knowing what to do with it, and that forms the political environment in which expert conversations have to actually succeed or fail. I think that level of fear is still mostly unaddressed in ways that would matter.
If Terror Management Theory is right, and if we can apply it to AI, then I think a lot of the current debate starts to make more sense. Someone whose career, community, or sense of purpose is tied to AI’s potential may not respond to existential-risk arguments by focusing on the argument alone. Instead, the focus can move to the people making the claim, and once identity is in the picture, disagreements are rarely just about facts anymore.
And yet, after all this, AI does not have its Hiroshima. No single image has done for this risk what the mushroom cloud did for nuclear weapons, or what the survivors’ testimonies did in the years that followed. There is no real catastrophe for people to point to, no equivalent of the firefighters’ widow describing a hospital room, because nothing has happened yet at the scale this fear is concerned with.
Calibrated Fear
Joanna Macy, writing about the nuclear and early climate years, saw something similar to what Lifton described. She argued that many people were living a kind of double life, carrying on with their ordinary routines while privately holding fears about the future that rarely entered conversation because there was little social space for them.
Something similar seems to be happening among people building AI today. Many researchers in major labs will say that they think there is a real chance their work could end badly. But very few organize their daily behavior as if that belief is central. I do not think this is hypocrisy. It may simply be the same thing Macy described, between private belief and public expression.
Seen this way, I think the issue is not simply a lack of information. It is that certain kinds of fear do not translate easily into shared social reality, even when they are taken seriously by the people who hold them. And if so, then I do not think this is the kind of problem more papers are going to solve.
Most of what has been written about AI risks so far is pure argumentation. Papers estimating timelines, essays comparing scenario A to scenario B, open letters with hundreds of signatures, and blog posts walking through failure modes step by step. Some of this work is genuinely excellent. But the comparisons drawn here are not arguments for more arguments. It is about things that cannot be readily manufactured by arguments alone. It is about testimony, about narrative, and about the collective cultural memory that gathers around them.
The testimony, the satire, and the narrative in the nuclear case did more than make people feel more intensely. It gave shape to the fear. It gave it details, actors, and consequences that could be identified, investigated, and acted on.
In the nuclear case, survivor testimony and cultural works helped, over time, to sustain the kind of attention that kept regulators, journalists, and legislators returning to a problem with no single dramatic resolution. Alexievich’s accounts of Chernobyl, for instance, don’t just talk about suffering in the abstract. They connect it to particular decisions, failures, and institutions. They do so in a form that is durable enough so that people who were not there can still read them as a record of what to refuse.
Calibrated fear is not released fear. It is fear that has been made specific enough to be actionable.
In practice, that might mean sustained public attention to a particular audit that was waived, a particular deployment decision that bypassed a safety review, or a particular exemption from oversight that someone put their name to. Not “AI is dangerous” as an ambient feeling, but this company, this choice, this consequence, as facts specific enough to refuse.
None of this diminishes the importance of arguments. The technical and governance work that actually reduces risk depends on it. But to sustain attention over time, we are missing the surrounding cultural layer that keeps attention specific, durable, and directed.
Two things complicate this:
The first is that nuclear fear got to be specific partly because it had Hiroshima and Chernobyl. That’s what testimony actually had material to work with. The risks people worry about most with AI right now may not have an equivalent, and I don’t think the answer is to wait for one. The diffuseness might not be a timing issue that a single bad event would eventually fix. It might be the actual shape of the risk. If so, the task isn’t finding our Chernobyl. It’s building specificity without the help of one.
The second complication is that even if the nuclear model is right and even if cultural works genuinely helped make nuclear fear specific enough to sustain political attention across time, the mechanism appears to have operated across decades. The Partial Test Ban Treaty was 1963. The ABM Treaty was 1972. The INF Treaty was 1987. So if the cultural soil that preceded those decisions was forming in the 1950s, then this mean that it took almost twenty to thirty years to become productive. It is very unclear whether AI risk operates on a timescale where that kind of slow cultivation is available. If it doesn’t, then the lesson from nuclear history may still be accurate and still arrive too late to be of use.
I don’t think these two complications cancel each other out. What they complicate is the timeline and the method. Whether specificity can be built without a catastrophe to solidify it, and whether it can be built fast enough to matter.
If that’s right, then the gap in AI discussions isn’t a shortage of feeling. It’s a shortage of address. Robot uprisings have an address. They point, however inaccurately, at something concrete enough to make a movie about. The risks that worry many AI researchers most, including the concentration of power in a handful of companies, the gradual erosion of human oversight, and the slow deterioration of our ability to verify what is true, do not.
Some AI risks, particularly the technical ones, may be addressable through the institutional route that made aviation safer without anyone outside the industry needing to feel anything in particular. But the risks this essay is most concerned with are the political and systemic ones, the ones that require external accountability to people who neither built the systems nor work inside the labs. For them, the nuclear model is the better template with all the slowness and uncertainty this implies.
This essay has not been arguing that people are not afraid enough, or that fear alone solves anything. This essay argues that fear of a systemic, slow-moving risk is politically and institutionally useful only when it is specific. When it attaches to specific decisions, specific institutions, specific failures that can be named and refused. Nuclear risk got there eventually, though it took decades and required a catastrophe to anchor it. AI risk has not gotten there yet, at least not for the diffuse structural risks that matter most. Whether it can, without waiting for the catastrophe to do the anchoring work, is the question the essay can identify clearly but cannot answer.
Right now, neither version exists in any developed form. We’re still in the part where Peter buys the bicycle wheels, informed, articulate about the danger, reaching for the nearest ordinary task instead of whatever specific thing might actually be done.
The Bulbs She'll Never See Bloom
Mary Holmes never stops planning her garden. That’s one of the last things we learn about her in On the Beach, after the radiation sickness has already reached the city, after the suicide pills have already been distributed at the local pharmacy like any other prescription. She keeps talking about bulbs. Not because she’s finally accepted what’s coming, and not because she’s still refusing to, but because there’s no longer a meaningful difference between those two states. The garden was never really about the garden.
We are, right now, somewhere in the middle of Peter’s afternoon. We understand the outline of the problem. We read the reports, follow the debates, argue about timelines and probabilities, and then go back to our ordinary lives. That doesn’t necessarily mean failure. Maybe it’s simply what people do when confronted with a threat they believe in abstractly but do not quite know how to incorporate into the ordinary routines of their lives, or how to respond to it.
History provides a record of what happened when the fears did get specific enough to act on. Nuclear fear did not remain forever at the level of abstraction. It became associated with certain places, events and failures over the years. It was concrete enough for people to build institutions around it.
Some AI risks may still be manageable via aviation’s quieter institutional route. But the political, systemic ones will not. Not when the institutions themselves are the ones concentrating the power.
I keep thinking about the woman in Pripyat watching the reactor burn from her balcony the night it happened, before anyone had a word for what was wrong. She said later that she had not known death could be so beautiful. The line holds both halves at once, the wonder and the wreckage, without choosing between them, as Mary Holmes holds her bulbs and her diagnosis in the same hand without choosing between them.
Somewhere, someone is buying bicycle wheels right now.
[This is the first essay I’ve ever published online, so I’d really value any feedback. I’m especially interested in critiques and perspectives that stress-test my reasoning. I’m also looking for writing tips if you have any. If you think I’ve missed something, I’d love to hear from you. Please do reach out!]
[Every argument in this essay is my own, developed through my reading and reflection. I used language models, specifically Claude and ChatGPT, only as editorial tools to critique my arguments and suggest improvements to the structure and wording.]
[This is cross-posted on EA Forum as well as Substack]
In Nevil Shute’s On the Beach, there is a scene where Peter Holmes, a naval officer living in suburban Melbourne, spends an afternoon looking for bicycle wheels. He needs them for a little trailer so his family can keep carrying groceries when the gas is gone. Meanwhile, a cloud of radioactive fallout is slowly moving south across the planet, killing all it touches. Peter knows what it is exactly. He’s studied the reports. He knows it will be in Melbourne in a matter of months and that it will kill him and his wife and their infant daughter. But he goes looking for bicycle wheels in the afternoon.
Mary, his wife, a few streets away, is planning a garden. She talks about bulbs she wants planted, a hedge she hopes will take root, and where a new flower bed should go. Peter tries to speak to her about what’s coming, but she gets annoyed and takes the conversation back to the garden. He lets her. There is no practical decision to be made in this part of the novel. All that remains is the ordinary business of living in the time they have left.
What is very interesting about this scene is how little the knowledge itself changes anything. Mary and Peter aren’t in denial. They read the same reports as everyone else does. At least in plain factual terms, they understand what’s coming and roughly when. Yet it changes almost nothing about how either of them spends the afternoon.
Something close to this is happening right now. People are spending their mornings asking a language model to draft an email (or maybe in my case, critique their writing) and their evenings reading papers by researchers inside some of these major AI labs who are estimating the odds that what they’re building could end up out of anyone’s control. Often this is the same person, in the same week. Nobody has stopped going to work. The wheels still need to be bought. The garden still needs to be planned. That’s not to say the comparison is meaningless. Peter and Mary had no options left. The cloud was coming regardless of what they did with their afternoon. But we still have options.
There is little evidence to think that more or less fear, by itself, determines whether anything meaningful gets done about these risks. I believe that fear, well directed, can help accelerate progress for obvious reasons. So maybe the question is why knowing about the danger still isn’t translating into behavior that looks proportionate to what people say they believe, even among people who are convinced that what happens next can still be changed. In this essay, I use history to help direct us in what could be the right direction.
I fully understand that AI will not follow the same path as nuclear weapons, for example, but at the same time, I believe that history is the only place where we can watch fear unfold from beginning to end. We can see how people make sense of it, how institutions respond to it, how it changes over time, and what happens when it is understood well or understood badly. The question this essay has been asking is never whether we should be afraid. The more interesting question, and arguably, the more important one, is what fear becomes once it enters public life, and whether it can be shaped into something useful before it shapes us instead.
Old Stories, New Machines
Societies do not usually invent entirely new fears when a new technology arrives. Instead, they reach for stories they already know and fit the new technology into them.
You can see this long before anyone had heard of atoms. Faust is just one example. The legendary figure was just an ordinary sixteenth-century German scholar whose story gradually turned into a cautionary tale about the dangers of pursuing forbidden knowledge. In the surviving version, Faust violates the boundaries of what he is supposed to know and ends up paying a terrible price for it. The details varied from telling to telling, but the message remained consistent: there are some kinds of knowledge that human beings seek at their own peril. A few centuries later, Mary Shelley gave that same fear a more modern shape in Victor Frankenstein. Frankenstein is not punished simply because he knows too much. He is punished because he creates something he cannot control. His mistake escapes into the world and begins affecting other people.
The details of the stories are different, but the pattern underlying them is the same. Someone acquires knowledge or power that was not meant to be theirs, and the consequences spread beyond their ability to manage them. By the time nuclear weapons appeared, that pattern was already familiar. The technology was novel. But the story people used to understand it was not.
The same appears to be true of thinking machines. Karel Čapek’s 1920 play R.U.R. introduced the word “robot,” from the Czech word for forced labor, and it also introduced a familiar storyline in which artificial workers created to serve humans eventually rebel and cause human extinction. Almost 20 years before the first general-purpose electronic computer was in existence, people were thinking about robot rebellion and the extinction of humans. A modern review in the New York Times referred to it as a “Czecho-Slovak Frankenstein,” which shows how quickly the story was mapped onto an even older and already familiar story. Decades later, the ship’s computer, HAL 9000 in 2001: A Space Odyssey, kills most of its crew after they attempt to shut it down. Skynet becomes self-aware in 1984 and turns on humanity. None of these stories was based on an intimate knowledge of how computers actually work. R.U.R. was written before the advent of digital computers. HAL and Skynet were way before the technical possibility of anything like an intentional machine. What they have in common is not technological prediction but a recurring narrative in which something created to serve humans eventually becomes a threat to human control.
This is not to say that the risks are not real. Most people who are worried about AI systems getting out of human control see that concern as a response to these powerful systems that are being built today, by huge labs with unprecedented acceleration in their capabilities. In part, it is. But the particular shape of worry, and I am talking about the image of betrayal and replacement, or that feeling that what was made will turn on its maker, this exact thing was there a long time before any of the relevant technology. People have often been right to worry about things, but for the wrong reasons. And that makes me wonder how much of what we feel about AI is actually a response to the systems themselves, and how much of it comes from an older story that has simply found a new form.
The Dread Map
American magazines in the mid-1990s ran a series of cover stories warning that flying had become dangerous. Commercial aviation, however, has killed fewer than 13,000 people across its entire history, going back to 1914. Meanwhile, cars kill roughly that many in the United States alone, most years. Yet, in the same years those magazine covers ran, surveys showed Americans listing air travel as one of their top dangers, much higher than car crashes.
Three decades later, and in 2024, around five billion passengers were flying on more than forty million flights worldwide, with the global accident rate remaining below one per million flights. Air travel is now one of the safest activities available to humans at scale. But somehow fear does not appear to be connected to these numbers at all. Surveys show that a substantial proportion of adults still report fear of flying, and their confidence in air travel can shift noticeably following recent incidents, even if the long-run safety trend remains constant. So, for example, after a cluster of high-profile crashes in early 2025, an AP-NORC poll found that the share of Americans describing air travel as “very” or “somewhat” safe fell from 71% to 64% in a single year.
This is the gap Paul Slovic spent much of his career trying to explain, and I don’t think it is just a gap in information. People who are more afraid of flying than driving usually know, at least abstractly, that driving is more dangerous.
Slovic and his colleagues asked people to rate dozens of hazards across a wide range of characteristics, including whether exposure was voluntary, how well the risk was understood by science, how catastrophic a single incident could be, whether the harm appeared immediately or only years later, and whether people could control it once it began. When they analyzed the data, most of the variation could be explained by just two dimensions. The first was dread, which described the feelings of a hazard being uncontrollable, catastrophic, and involuntary. The second was the unknown, which measured how invisible, unfamiliar, and delayed its effects seemed. A hazard’s position on this map was what predicted how much regulation people wanted and how loudly they wanted it stopped. As you may have already noticed, it does not depend on how many deaths a hazard contributed to. Nuclear weapons, as you may also have assumed, sit at the far edge of both axes at once, which I think is part of why they generated a fear wildly disproportionate to how often they’ve actually been used.
Cass Sunstein took this even further. He called this phenomenon “probability neglect,” which describes an outcome that is emotionally vivid enough that people stop adjusting their concern based on how likely it is. One experiment illustrates this idea very well. Participants were asked how much they would pay to avoid a short, painful electric shock. When the probability of the shock increased from 1% to 99%, the amount they were willing to pay rose only slightly. But when the same range of probabilities was attached to a cash penalty instead of a shock, willingness to pay rose and fell much more in line with the actual odds. In both cases, the numbers were identical. What changed was the emotional weight attached to them. Sunstein did a different version of this with cancer risk from arsenic in drinking water, and the pattern was the same. A tenfold increase in the actual risk barely changed what people would pay to fix it. But after making the description of the cancer sound more frightening, without changing the odds at all, it moved them almost as much as the tenfold increase in real danger had.
This also works the other way around. The same mechanism that causes people to overreact to a vivid, rare threat also causes them to underreact to a duller, more common one. Sunstein found that when people were told the odds of dying on a given car trip were one in four million, people would skip their seatbelts. Maybe it was because the number was so small that people simply stopped thinking about the likelihood of this happening at all. But why is this not the case for air travel we have just explored above? A society can be terrified about one thing and nearly numb to another that kills far more people, and yet the intensity of that fear tells us almost nothing about the actual danger.
Culture shapes our perceptions of risk in ways that individual psychology alone cannot explain. Mary Douglas and Aaron Wildavsky were interested in understanding the reasons. They wanted to know why sometimes entire communities fear different things. They found that risk selection also depends on culture, not just individual psychology. Different social groups, hierarchical institutions, tight-knit egalitarian movements, and market-driven individualists are driven by their own structure to see certain dangers as self-evident and others as nearly invisible. This matters because the people most concerned about an AI system getting outside anyone’s control and the people who categorize that as science fiction are not, in most cases, working from different facts. They may just be embedded in different social worlds that contribute to their perception of what is danger and what is not.
So, putting all of this together. Statistics do not seem to promote fear, and little difference is made when they do. Fear is created and promoted by the feelings of people. And once something is vivid enough, it crowds out attention to whatever is quieter, even if it was more probable. This is the exact mechanism that lets an extremely rare plane crash dominate magazine covers while far more common driving deaths go ignored entirely. Even which version of this story gets told depends on which community you belong to.
The Second Catastrophe
On the morning of August 6, 1945, a bomb exploded with the force of about twenty thousand tons of TNT eighteen hundred feet above the city of Hiroshima. Within a half-mile radius, more than nine in ten people standing outdoors and unshielded were dead within the hour. Sixty thousand of the city’s ninety thousand buildings were gone. Estimates of the dead range from the high tens of thousands to the low hundreds of thousands, and nobody has ever settled on an exact number.
None of the cultural machinery described above (such as the mad scientist, or the secret that’s too dangerous to know, or even the created thing that somehow turns on its maker) was built to hold something like this. Those stories were myths about a danger that hadn’t happened yet. But Hiroshima had happened. As it turns out, the distance between an abstract dread that has been circulating for decades and the real, dated, photographed, named catastrophe is enormously significant.
Robert Lifton, who interviewed dozens of survivors years later, found that the devastation was so extreme that it felt completely unreal. Survivors often described a similar sequence. They saw a flash, then a silence that seemed too long, then a sound like distant thunder, and then a world that no longer looked familiar in any way. People who had been three miles away were as certain the bomb had landed beside them as those who had actually been near its center. Many remembered wandering through the ruins with no destination in mind, simply moving forward in something closer to slow motion than panic. One word appeared again and again in the interviews. It was a term that meant something like “without self” or “without a center.”
What Lifton’s interviews also showed was a defense mechanism that was so fast-acting that most survivors didn’t recognize it was happening to them until much later. One survivor, for instance, was a man assigned to mass cremations who described handling corpses the way you’d handle freight, with almost no emotion at all. Or a doctor who kept a diary through the weeks after the bombing wrote that he had come to see death as ordinary and considered a family fortunate if it had lost only two members. I do not see this as callousness in any normal sense. I’m sure that these people still cared deeply about what was happening around them. But it was the amount of suffering they were exposed to that altered their capacity to respond. Lifton later gave this a name. He termed it “psychic numbing”.
Years later, the same man who had handled the corpses without feeling anything described how horrified he was, in retrospect, at his own businesslike behavior. The numbing was not simply a survival mechanism of the event. It was, in its own way, a small death running parallel to the larger one, and people carried both through the rest of their lives.
All of this is about the blast itself, the part of the catastrophe that occurred in that single instant, and was at least, in the end, visible and countable. The second catastrophe took longer to even be recognized as a catastrophe. It is the one that maps most directly onto what worries people about AI today.
Within days, people who had walked away from the blast without any visible injury started getting sick. The symptoms often begin with nausea, followed by bleeding gums, then patchy hair loss, and then, in many instances, death, which can happen even weeks after the person appeared to have recovered. There was no way to tell, just by looking at someone, whether they were going to be fine or whether they were already dying. For weeks, the city doctors had no name for this thing. Tokyo radio reports first said that something was killing relief workers who were sent to help. An American science writer with a thin connection to the bomb program told the press that Hiroshima would be lethal to enter for seventy years, and that the radiation would destroy red blood cells and cause leukemia in anyone exposed. This was then withdrawn under pressure within days, but a withdrawal, as we may already know, rarely travels as fast or as far as the claim it’s correcting. As a result, rumors began to spread that nothing would ever grow in the city again and that no one could live there for decades. People said that the dead were too radioactive to be buried normally. Though none of it was accurate. The window of serious residual radiation danger was measured in weeks. Though numbers didn’t really matter. People examined themselves, and each other, for the small purple spots that indicated internal bleeding, a habit one doctor in the city described, in his own diary, as something close to a citywide epidemic of suspicion.
Before August 1945, the fear was mainly of the explosion itself, its size, its fire, and that one most overwhelming moment. The explosion did happen, and it was every bit as catastrophic as expected, but it was over in seconds. But the invisible thing that was with no smell, no sound, no clear timeline was what lasted, and it came to define how people related to their own bodies and their city for years. It was something you could not see coming and could not be sure you had escaped. The bomb people had feared, and the threat that actually governed their lives afterward, were not, in the end, the same thing at all.
The institutions that were supposed to deal with this had a different problem. It is what to tell people about it. There’s a film, shown in American classrooms throughout the 1950s, in which a cartoon turtle named Bert tells children what to do if they see a flash brighter than the sun. You duck down, cover your head with your arms, and get under your desk. Schoolchildren practiced this for years, some of them really believing, because the film told them so in a calm and reassuring voice, that these steps would protect them if a nuclear weapon went off nearby. The historian Guy Oakes argued that the real purpose of civil defense was not simply to teach survival techniques. It was to preserve the sense that the situation remained orderly, manageable, and under control. It was to show that someone in authority had a plan. Whether those plans would have been enough in a real attack was, in many ways, a secondary question.
What is interesting is that people seemed to feel this, even if they could not always explain why. Years later, one civil defense advocate ran an informal experiment in Baltimore and found that almost no one could recall seeing the shelter signs posted throughout their own neighborhoods, despite passing them regularly. The signs were visible, but people had learned to look past them. And perhaps that’s the lesson. When institutions focus on maintaining the appearance of safety rather than confronting the danger itself, they do not always create reassurance. Sometimes they simply teach people to stop paying attention.
Three Ways Fear Goes Wrong
How does a collective fear held so strongly end up generating almost no proportionate response? There are three distinct lines of research that converge on more or less the same conclusion.
The first draws from Terror Management Theory, built out of the work of psychologist Ernest Becker on mortality and tested across hundreds of experiments since, though several of its core findings have faced replication challenges in recent years and remain contested within the field. It says that reminding people of a threat does not necessarily make them think more clearly about threats. In many cases, it does the opposite. It pushes them deeper into the beliefs and identities they already hold. In one of the experiments testing this theory, judges asked to set bail for a minor offense set it roughly nine times higher after being asked two questions about their own mortality first, with nothing else about the case changed. The effect actually has nothing to do with the specific threat. It’s on whatever a person’s identity is already built on. This matters outside the lab. It means that those who are most likely to reject an uncomfortable argument are not necessarily those who have thought least about it. It could just be that they are the ones whose argument threatens something that they already need to believe in.
There’s a related finding, sometimes described as the distinction between near and far threats, that matters even more here. A mugger on a dark street causes a quick, practical, embodied fear that makes you walk faster or cross the road. But for risks that are real but abstract and years away, things like climate change, for example, or an AI system that might slowly erode human oversight over the coming decade, these usually don’t trigger that same machinery at all, even when the actual stakes involved are far greater. It seems the mind has one fear system for things close enough to act on immediately, and a much weaker, more talk-yourself-out-of-it system for things that are real but distant. This is part of how a society can think that a threat is serious and yet do almost nothing about it.
Glassner’s work fills in what the first two frameworks leave out. It says that there is always someone who profits from fear pointing where it points. Glassner spent years tracing the way American fear in the 1990s got attached overwhelmingly to rare, vivid dangers while statistically much larger ones went almost as nothing. Drunk driving, for example, killed roughly eighty times more people a year than road rage did, but road rage got the press coverage, partly because it made for better TV shows, and partly because blaming an individual’s temper is much easier and less politically costly than asking why there are so many guns in cars. I think something similar happens whenever a vivid, cinematic AI scenario gets more attention than a duller, more probable one. It isn’t necessarily that anyone’s lying. It’s that the vivid version is easier to fund, easier to cover, and easier to build a career or a movement around than the duller one will ever be.
Stanley Cohen’s work on moral panics offers a useful way of checking whether a fear has begun to drift away from the thing it is supposedly about. He suggests looking out for a few warning signs. Things like: Is the public reaction disproportionate to the evidence of harm? Is the coverage moving faster than the facts can keep up with? And is there a clearly identifiable target that ends up carrying anxieties that are really about something much broader? Applied to the AI context, the first two are not hard to find. There is plenty of amplification in the discourse, and at times plenty of disproportion as well. The third question is more complicated. Public criticism often gets concentrated on a small number of companies, such as OpenAI or Anthropic. In that sense, they do seem to be the kind of visible target Cohen had in mind.
But there is also an important distinction, too. The groups Cohen wrote about were often much weaker than the institutions that criticized them. AI companies are not. They have a lot of money, a lot of political influence, and a lot of public platforms. But that does not make the criticism incorrect. It simply means the situation does not fit neatly into Cohen’s framework.
So let’s put these three frameworks together and see what they suggest. Terror Management Theory explains what fear can do to a person. It shows that under threat, people often double down on their existing identities, beliefs, and groups. Glassner and Cohen are more concerned with what goes outside the individual. They show how fears can be magnified, minimized, redirected, or attached to particular targets as they move through institutions, media, and public discussion.
All of this was true for nuclear weapons, and all of it is still true for AI today. The question is, what historically actually made this kind of distorted, identity-driven, institutionally-managed fear into something specific enough to act on?
It would be dishonest not to note that these frameworks cut in both directions. If cultural community influences which dangers seem real, then the community most worried about AI risks is itself a community with its own structural priors. The same probability-neglect mechanism that makes cinematic scenarios crowd out diffuse ones could operate among believers as well as skeptics, producing distortion upward rather than downward.
The Three Techniques
History offers three answers to that question, and none of them are arguments. An institution may sincerely try to manage fear and still end up only managing its appearance, as the civil defense programs described earlier suggest. But what institutions cannot do, at least not by themselves, is transform diffuse dread into something that people actually understand and know how to act on. That work, in the nuclear case, came from somewhere else entirely.
Dr. Strangelove opens with a joke that only makes sense if everyone in the room already knows how serious the subject matter is. When a rogue general is convinced that fluoridated water is part of a Communist plot to sap American “vital fluids,” he launches an unauthorized nuclear strike. The president calls the Joint Chiefs into a room of bleak, giant maps, and within minutes, two of his most senior officers are wrestling on the floor. That’s when the president delivers the line that has outlived the rest of the film by half a century: “Gentlemen, you can’t fight in here. This is the War Room.” The film’s actual subject, the systems and the men behind them that could genuinely end the world by accident, never once stops being terrifying for the ninety minutes it spends being funny.
What satire does, when it is aimed at something this big, is less to defuse the threat than to give it a shape small enough to hold for a moment, a shape with a name and a face and a particular, identifiable absurdity, instead of the shapeless, statistical dread of “nuclear war” in the abstract. You can’t really organize a feeling against “the possibility of mutual annihilation.” You can organize a feeling against a general who thinks his vital fluids are being attacked, or a wheelchair-bound ex-Nazi scientist who can’t stop his arm from saluting, or a grinning cowboy who rides a bomb down to earth like a rodeo bronco. The film ends with mushroom clouds blooming across the screen to a nostalgic wartime love song, and audiences reportedly left in tears rather than laughing, which is its own evidence that something real had been delivered underneath the joke. I don’t think satire made anyone less afraid of nuclear war. It gave the fear a villain, a failure mode, a specific moment when one absurd decision cascades into catastrophe, something concrete enough to actually be mad about.
Nevil Shute did almost the opposite and came to something close to the same point by a different route. On the Beach has no jokes, no villains, no single moment of decision. There is a submarine commander who decides to sink his own ship rather than abandon his post, because some idea of duty still makes sense even now. But the novel will not develop into anything bigger than that. It remains, relentlessly, at the level of the ordinary. None of this changes the outcome. What it changes is the reader’s relationship to the statistic. No one can hold a sentence like “Several hundred million people will die of radiation sickness” in their head for more than a second. A specific woman planting specific bulbs, in a specific garden she has already decided not to think too hard about, is something a reader can carry around afterward. Humanization didn’t make the threat smaller so much as make it specific enough to actually grieve, which is an entirely different operation from simply being informed about it.
The third technique does not dramatize or humanize at all. It simply records. After Chernobyl, Svetlana Alexievich spent years collecting testimony from those who had actually lived inside the disaster, including liquidators, widows, evacuees, a woman who watched her husband die of radiation sickness over two weeks in a Moscow hospital, and who was told not to touch him. One woman, who is an environmental inspector who had spent the cleanup watching colleagues and officials each come up with their own little, reasonable-sounding justifications for looking away, arrived by the end of her testimony at something close to a thesis for the whole collection. The worst things in a system rarely show up as one dramatic, villainous act. They come to us through ordinary people, one decision at a time, each one defensible on its own. That insight is testimony as a technique. It doesn’t turn the catastrophe into a villain’s plot, nor condense it into the story of one sympathetic family. It lays out, in plain, specific, often bureaucratic detail, just how an entire system of people, none of whom are monsters, produced something monstrous. The reader is not given an emotion to release. They are given a mechanism: who signed what, who looked away from what, who pulled the contaminated milk off the shelves and who did not, who was told the truth and who was lied to, and exactly how.
None of the three made people feel less afraid. All three made the fear sharp enough to land somewhere. Not necessarily as immediate policy action, but as cultural memory specific enough to hold attention across decades, and durable enough to carry forward long after any single political moment had passed.
AI doesn’t have any of this yet. No Strangelove to give the risk a face absurd enough to be angry at. There is no On the Beach where the “loss of human oversight” is one person losing one thing. There’s no Alexievich sitting across from the people closest to whatever the first real AI failure turns out to be, taking notes on who knew what and when. What AI risk has instead is a few hundred technical papers, a handful of essays from researchers inside the major labs, and continuous arguments on social media. It is the sort of speech meant to persuade and to reply, not to make a threat specific enough to act upon. But none of it does what satire, humanization, or testimony each managed to do in their own way, which is take something too large and too abstract to hold and hand people back something they could actually point at.
The Social Dilemma, for instance, comes closest to testimony as it names specific companies, specific design choices, and specific consequences, and the harms it describes are real and alarming. But it named only specific risks. The Social Dilemma builds accountability for attention manipulation and addictive design. What remains unaddressed is a different set of failures entirely. Things like the concentration of decision-making power in a handful of unelected companies, or the erosion of the oversight mechanisms researchers mean when they talk about systemic risk, or the slow degradation of what anyone outside these labs can even verify as true. Those risks don’t yet have their Social Dilemma. Maybe one reason is that they are structurally harder to make specific, so they don’t arrive in a single dramatic design choice, and they don’t have a single sympathetic victim whose story can be told.
It is also worth noting that none of this can be manufactured on demand. Dr. Strangelove was born from Kubrick’s pre-existing obsession with the absurd and the right collaboration. Alexievich spent years in Chernobyl because it was the subject before it became anyone’s priority. The works that eventually made nuclear fear specific were not commissioned by frightened researchers. They were produced by artists who were listening to something that frightened them. Which makes the absence harder to address than simply deciding to address it.
Nuclear or Aviation?
Before conceding the importance of this absence, it is worth asking whether public cultural engagement with a dangerous technology is actually what makes it safer. Aviation is a direct challenge.
Commercial aviation has gotten dramatically safer over the last several decades. Arguably, the greatest safety improvement in the history of any transportation technology, and it happened entirely without a mass cultural reckoning. No film did for plane crashes what Dr. Strangelove did for nuclear war. No novelist gave the world an aviation equivalent of On the Beach. What drove the improvement was simply more institutional. After a series of catastrophic accidents in the 1970s, the industry created blameless incident reporting: a culture in which pilots and crew could report near-misses, mistakes, and safety concerns without fear of punishment or career consequences. Before this was around, errors were buried because admitting them cost careers. After that, near-misses became data. So black boxes helped with that. Training in crew resource management was developed directly from studies of cockpit interactions that broke down in particular accidents. None of this required the public to feel anything in particular. It required insiders to build a professional culture that viewed failure as information that helps make the upcoming journeys safer.
There are people trying to build something similar inside AI labs right now. Things such as the red-teaming cultures, internal safety reviews, and responsible scaling policies. If the risks that matter most in AI turn out to be technical enough that insider professional culture can address them without public engagement, then the nuclear model is the wrong template, and this essay’s concern is misplaced. But I think this is not the case.
There are two reasons AI is more likely to need the nuclear model than the aviation one.
Aviation’s risks were mechanical and the fixes were technical. An aircraft crashes because of a specific mechanical failure, a navigation error, or a breakdown in cockpit communication. All of these are problems that engineers and professional culture can solve without anyone outside the industry needing to understand or care about them. The risks that most concern AI researchers are not like this. Power concentration in a handful of companies, erosion of meaningful human oversight, or the slow degradation of what anyone can verify as true, all of these are political and systemic problems. They require regulatory pressure, legislative attention, and an ongoing public accountability that only exists when enough people outside the industry care sufficiently to ask for it. You cannot fix the political economy of AI development with blameless incident reporting inside the labs. It’s the labs that are the ones accumulating the power.
The near-misses of aviation were also readable to the insiders who could fix them. Pilots and crew could see failures and explain them well enough for engineers to act on. Many of the AI risks that people are most worried about are not even clearly readable to insiders. This means insider professional culture alone, however healthy, cannot by itself catch the failures that matter most.
No Hiroshima Yet
The difference is that nuclear fear had an event. A catastrophe that was dated and photographed and named, with survivors who could be interviewed and a death toll that could be estimated, however roughly, in the end. AI fear doesn’t have that, and maybe it never will in quite that same form. Instead, a lot of the field’s imagination has been driven by thought experiments, by predictions, and by a handful of influential books and movies.
The closest thing AI risk discourse has found to an anchoring text is instructive about both what this can and cannot do. Nick Bostrom’s Superintelligence (2014) is one of the most influential books written about AI risk. Its central argument is that an AI system that’s sufficiently capable would not have to be malicious to become dangerous. If it had a goal and was powerful enough, it might resist attempts to shut it down, resist having its goals changed, and seek out additional resources to fulfill whatever objective it had been given. One possible version of this dynamic is the “treacherous turn,” in which a system behaves cooperatively so long as its interests are served, but then changes its behavior once it no longer does.
Intelligent machines had been a cause for concern for decades. What Bostrom did was to give those concerns a more coherent structure. In that sense, the book played a role similar to the one Spencer Weart describes in his history of nuclear fear. It gave an existing anxiety a clearer shape, and once it had that shape, it became much easier to communicate, debate, and spread.
Many of the AI risks that researchers are most concerned about do not have those same advantages. It’s hard to reduce things like gradual concentration of power in a small number of companies, or the slow erosion of meaningful human oversight, or even the possibility that AI systems make it harder to distinguish truth from falsehood, to a single dramatic image. That does not, and should not, make them more important or less important than existential risks. It just simply makes them harder to picture, and therefore easier to overlook.
In recent years, AI fear has taken on a real specificity at the level of expert and policy attention. Regulatory frameworks like the EU AI Act directly address risks such as power concentration and systemic risk. Geoffrey Hinton’s public departure from Google drew mainstream coverage that was focused on oversight erosion. Senate hearings have forced AI executives to testify about the very institutional failures that researchers worry about. Elite discussion, in short, has found some address. What I am describing is the background cultural anxiety that most people carry around, not knowing what to do with it, and that forms the political environment in which expert conversations have to actually succeed or fail. I think that level of fear is still mostly unaddressed in ways that would matter.
If Terror Management Theory is right, and if we can apply it to AI, then I think a lot of the current debate starts to make more sense. Someone whose career, community, or sense of purpose is tied to AI’s potential may not respond to existential-risk arguments by focusing on the argument alone. Instead, the focus can move to the people making the claim, and once identity is in the picture, disagreements are rarely just about facts anymore.
And yet, after all this, AI does not have its Hiroshima. No single image has done for this risk what the mushroom cloud did for nuclear weapons, or what the survivors’ testimonies did in the years that followed. There is no real catastrophe for people to point to, no equivalent of the firefighters’ widow describing a hospital room, because nothing has happened yet at the scale this fear is concerned with.
Calibrated Fear
Joanna Macy, writing about the nuclear and early climate years, saw something similar to what Lifton described. She argued that many people were living a kind of double life, carrying on with their ordinary routines while privately holding fears about the future that rarely entered conversation because there was little social space for them.
Something similar seems to be happening among people building AI today. Many researchers in major labs will say that they think there is a real chance their work could end badly. But very few organize their daily behavior as if that belief is central. I do not think this is hypocrisy. It may simply be the same thing Macy described, between private belief and public expression.
Seen this way, I think the issue is not simply a lack of information. It is that certain kinds of fear do not translate easily into shared social reality, even when they are taken seriously by the people who hold them. And if so, then I do not think this is the kind of problem more papers are going to solve.
Most of what has been written about AI risks so far is pure argumentation. Papers estimating timelines, essays comparing scenario A to scenario B, open letters with hundreds of signatures, and blog posts walking through failure modes step by step. Some of this work is genuinely excellent. But the comparisons drawn here are not arguments for more arguments. It is about things that cannot be readily manufactured by arguments alone. It is about testimony, about narrative, and about the collective cultural memory that gathers around them.
The testimony, the satire, and the narrative in the nuclear case did more than make people feel more intensely. It gave shape to the fear. It gave it details, actors, and consequences that could be identified, investigated, and acted on.
In the nuclear case, survivor testimony and cultural works helped, over time, to sustain the kind of attention that kept regulators, journalists, and legislators returning to a problem with no single dramatic resolution. Alexievich’s accounts of Chernobyl, for instance, don’t just talk about suffering in the abstract. They connect it to particular decisions, failures, and institutions. They do so in a form that is durable enough so that people who were not there can still read them as a record of what to refuse.
Calibrated fear is not released fear. It is fear that has been made specific enough to be actionable.
In practice, that might mean sustained public attention to a particular audit that was waived, a particular deployment decision that bypassed a safety review, or a particular exemption from oversight that someone put their name to. Not “AI is dangerous” as an ambient feeling, but this company, this choice, this consequence, as facts specific enough to refuse.
None of this diminishes the importance of arguments. The technical and governance work that actually reduces risk depends on it. But to sustain attention over time, we are missing the surrounding cultural layer that keeps attention specific, durable, and directed.
Two things complicate this:
The first is that nuclear fear got to be specific partly because it had Hiroshima and Chernobyl. That’s what testimony actually had material to work with. The risks people worry about most with AI right now may not have an equivalent, and I don’t think the answer is to wait for one. The diffuseness might not be a timing issue that a single bad event would eventually fix. It might be the actual shape of the risk. If so, the task isn’t finding our Chernobyl. It’s building specificity without the help of one.
The second complication is that even if the nuclear model is right and even if cultural works genuinely helped make nuclear fear specific enough to sustain political attention across time, the mechanism appears to have operated across decades. The Partial Test Ban Treaty was 1963. The ABM Treaty was 1972. The INF Treaty was 1987. So if the cultural soil that preceded those decisions was forming in the 1950s, then this mean that it took almost twenty to thirty years to become productive. It is very unclear whether AI risk operates on a timescale where that kind of slow cultivation is available. If it doesn’t, then the lesson from nuclear history may still be accurate and still arrive too late to be of use.
I don’t think these two complications cancel each other out. What they complicate is the timeline and the method. Whether specificity can be built without a catastrophe to solidify it, and whether it can be built fast enough to matter.
If that’s right, then the gap in AI discussions isn’t a shortage of feeling. It’s a shortage of address. Robot uprisings have an address. They point, however inaccurately, at something concrete enough to make a movie about. The risks that worry many AI researchers most, including the concentration of power in a handful of companies, the gradual erosion of human oversight, and the slow deterioration of our ability to verify what is true, do not.
Some AI risks, particularly the technical ones, may be addressable through the institutional route that made aviation safer without anyone outside the industry needing to feel anything in particular. But the risks this essay is most concerned with are the political and systemic ones, the ones that require external accountability to people who neither built the systems nor work inside the labs. For them, the nuclear model is the better template with all the slowness and uncertainty this implies.
This essay has not been arguing that people are not afraid enough, or that fear alone solves anything. This essay argues that fear of a systemic, slow-moving risk is politically and institutionally useful only when it is specific. When it attaches to specific decisions, specific institutions, specific failures that can be named and refused. Nuclear risk got there eventually, though it took decades and required a catastrophe to anchor it. AI risk has not gotten there yet, at least not for the diffuse structural risks that matter most. Whether it can, without waiting for the catastrophe to do the anchoring work, is the question the essay can identify clearly but cannot answer.
Right now, neither version exists in any developed form. We’re still in the part where Peter buys the bicycle wheels, informed, articulate about the danger, reaching for the nearest ordinary task instead of whatever specific thing might actually be done.
The Bulbs She'll Never See Bloom
Mary Holmes never stops planning her garden. That’s one of the last things we learn about her in On the Beach, after the radiation sickness has already reached the city, after the suicide pills have already been distributed at the local pharmacy like any other prescription. She keeps talking about bulbs. Not because she’s finally accepted what’s coming, and not because she’s still refusing to, but because there’s no longer a meaningful difference between those two states. The garden was never really about the garden.
We are, right now, somewhere in the middle of Peter’s afternoon. We understand the outline of the problem. We read the reports, follow the debates, argue about timelines and probabilities, and then go back to our ordinary lives. That doesn’t necessarily mean failure. Maybe it’s simply what people do when confronted with a threat they believe in abstractly but do not quite know how to incorporate into the ordinary routines of their lives, or how to respond to it.
History provides a record of what happened when the fears did get specific enough to act on. Nuclear fear did not remain forever at the level of abstraction. It became associated with certain places, events and failures over the years. It was concrete enough for people to build institutions around it.
Some AI risks may still be manageable via aviation’s quieter institutional route. But the political, systemic ones will not. Not when the institutions themselves are the ones concentrating the power.
I keep thinking about the woman in Pripyat watching the reactor burn from her balcony the night it happened, before anyone had a word for what was wrong. She said later that she had not known death could be so beautiful. The line holds both halves at once, the wonder and the wreckage, without choosing between them, as Mary Holmes holds her bulbs and her diagnosis in the same hand without choosing between them.
Somewhere, someone is buying bicycle wheels right now.
[This is the first essay I’ve ever published online, so I’d really value any feedback. I’m especially interested in critiques and perspectives that stress-test my reasoning. I’m also looking for writing tips if you have any. If you think I’ve missed something, I’d love to hear from you. Please do reach out!]