For the people of the United States and the other countries that were spared the worst of the fighting on their own soil during the Second World War, the Cold War was perhaps the period when the phrase "sadder but wiser" began to apply to a far wider array of ordinary people. Almost everyone was aware, in some dim and terrible way, of the devastation the American bombings of Hiroshima and Nagasaki had visited on Japan in 1945. But for most, this was not something that could happen to them. For a time, the United States was the only country with a nuclear arsenal. And wars, in any case, were for the most part something that happened in far-away places, posing no real danger of striking home.

That reprieve lasted almost exactly four years. In August 1949, enabled by physicist Klaus Fuchs and a cohort of American communist sympathisers, the Soviet Union detonated its first atomic device. The monopoly was broken. What followed was an escalation of tests and technologies on both sides that produced, almost as a by-product, the first great wave of messaging, sponsored and distributed by governments, about how nuclear weapons work, how a war might unfold, and what might be the consequences for human life.

The messaging was, by turns, reassuring and terrifying. American schoolchildren were taught to "duck and cover" by an animated turtle called Bert, as though a school desk might be of some use against the destructive capacity of a nuclear weapon. Families were instructed to stock up on tinned food and to consider building a fallout shelter in the back garden. Much like the advice given by the British government a generation later to "Protect and Survive", with its instructions on whitewashing windows and the identification of a "fall-out room", this was almost entirely designed to give citizens some sense of control and autonomy, a feeling that they could do something to protect themselves and others, rather than something that was bleakly honest. At once, a nation of some 150 million people, many of whom would happily have remained blissfully ignorant, was subjected to the avalanche of a strange and frightening new vocabulary: radioactive fallout, the splitting of the atom, fission and fusion, megatonnage, the electromagnetic pulse. The future had arrived, and it was horrible, and it seemed inescapable.

An Age That Forgot to End

As far as I or most anyone else growing up in Western Europe during the late 20th & early 21st century could tell, the Nuclear Age had ended. The Berlin Wall had come down, the Soviet Union had evanesced into a campy Bond villain, and a generation grew up at least half-believing that the bomb had been filed away with the other horrors of the twentieth century. That belief now looks like another short-lived fairytale to confine to the history books; an optimistic, perhaps even hubristic fallacy. We are living through a second age of nuclear proliferation, and it is in many ways far more dangerous than the first for at least three reasons: there are more nuclear-armed powers; the geopolitical landscape is increasingly multipolar; and the grim etiquette that two superpowers had decades to negotiate has been eroded over time.

The signs are not subtle. At the start of this year, the Bulletin of the Atomic Scientists set its Doomsday Clock at 85 seconds to midnight, the closest to catastrophe it has stood in its history, citing an outright arms race in which China is rapidly expanding its warhead numbers and the United States and Russia are modernising their delivery systems. As the Bulletin published that statement, New START - the last major treaty limiting the strategic arsenals of the two largest nuclear powers - was set to expire, ending nearly six decades of constraint. In a piece in Nature last year, science journalist Alexandra Witze summarised this shift bluntly: nuclear deterrence "is no longer a two-player game", and emerging technologies are further destabilising an already precarious status quo.

Nor is this an abstract, arithmetical concern about stockpiles. In the space of a few months, the world watched three regional conflicts involving nuclear-armed states threaten to boil over. In May last year, India and Pakistan fought their most serious direct engagement in decades. Then in June, Israel and the United States struck Iranian nuclear facilities. And all of this unfolded against the backdrop of Russia's continuing war on Ukraine; a war punctuated by repeated allusions from Moscow (though frequently ignored as hollow by European diplomats) regarding the use of nuclear weapons. Nine countries now hold these weapons (Israel's "strategic ambiguity" has long been far more "strategic" than "ambiguous"), and the guardrails that once contained them are corroding.

It is into this environment of heightened conflict, expanding arsenals, decaying treaties that a genuinely new variable has been introduced. One that the men who dreamt up Bert the Turtle could not have imagined.

Ignorance, as a Salve and a Curse

A colleague of mine recently spoke with a prospective client about how we might use AI & data analytics in the roughly $50 million trucking business he owned. Partway through the conversation, the prospect revealed that he did not use artificial intelligence. For anything. Not only that, but beyond a headline he had half-read, a topic that had come up once at dinner, he was barely aware of it.

Fewer and fewer of these rare specimens remain. For the overwhelming majority, AI is now pervasive. It is not merely in every headline, every podcast, and every newsletter; it is force-fed to us through nearly every piece of software we touch, every search engine, every application, every website. Welcome or otherwise, it has for many people become the first port of call for information, the thing we ask before we ask a human, the summary we read instead of the source.

This, in the view of the author, is the most immediate and widest-reaching threat that AI poses in this second age of expanding nuclear arsenals. Not, or rather, not yet, the prospect of a black-box algorithm with its digital 'finger' on the button (though that danger is real and growing), but something more insidious and already here: the pollution of the entire information environment in which nuclear weapons, nuclear war, and nuclear energy are discussed, understood and, in a crisis, acted upon.

AI has already been deployed by individuals, organisations, and nation states to disinform millions. Deepfake videos of politicians saying and doing things that never happened. Mobs mobilised, and in some cases roused to violence, by wholly synthetic content. The industrialised automation of propaganda: brainstormed, generated, and distributed at scale with the press of a few buttons. There is a small silver lining here: the sheer scale of this means that people with low IT literacy who may never before have questioned the source or truth of online content are perhaps slightly more likely to recognise something as having been AI-generated (example: my wife's grandmother correctly identified that a video of a dancing dog was "probably AI"). This is a very small consolation, however, because as the technology continues to improve and fakes become harder to spot, they are more likely to take hold. The energy required to disprove any one piece of it is (as Brandolini's Law has it) an order of magnitude greater than the energy required to create it. And the platforms that host this material are frequently both unwilling to remove it, because outrage and confrontation drive the engagement on which their revenue depends, and unable to act before it has already spread too far. To imagine that this dynamic will somehow spare the single most consequential subject in human affairs is beyond naïve.

The Liar's Dividend

The core of the problem is an asymmetry. Falsehood is cheap; correction is dear. During any crisis, "falsehoods tend to travel farther and faster than corrections" observed Ethan Beaty in the Bulletin last September. He was drawing on a much-cited 2018 study by researchers at the Massachusetts Institute of Technology, which found that false news on X (then called Twitter) spread "farther, faster, deeper, and more broadly" than the truth; reaching people up to six times faster and being retweeted 70 per cent more often than accurate reports. Generative AI, which arrived in consumer-friendly form in 2022, has simply industrialised the manufacture of the falsehoods in question.

There is a more subtle cost, too. Legal scholars call it the "liar's dividend": once the public knows that convincing fakes exist, people become more ready to dismiss genuine evidence as fabricated, and to hesitate when a real warning arrives. In an ordinary emergency, that hesitation costs lives. In a nuclear one, it would cost exponentially more. Brian Cook, a policy research analyst for the Los Alamos National Laboratory, notes in a piece for the Centre for Strategic and International Studies that the difficulty is compounded by scale and anonymity: bot networks can inflate the apparent popularity of a claim, and large language models can generate limitless plausible text. A University of Notre Dame experiment he cites found that 58% of the time, participants could not reliably tell whether they were arguing with a human or a machine.

Nuclear weapons happen to be an almost perfectly designed target for this machinery. The subject is highly technical, largely classified, and intensely emotive; a combination that leaves the public unusually dependent on trusted intermediaries and therefore particularly vulnerable when those intermediaries are impersonated. Cook points to research showing that people are more likely to believe a claim simply because they have encountered it repeatedly, the so-called illusory-truth effect, and that heightened emotional states increase belief in fake news specifically, while leaving belief in true news unchanged. Fear, in other words, is not incidental to nuclear disinformation. It is the mechanism.

South Asia: A Grim Rehearsal

To see how these forces combine, look at the clash between India and Pakistan in May 2025. In the hours after Indian strikes on Pakistani airbases, social media became "the epicentre of a digital counter-offensive." AI-generated satellite images, deepfaked footage of explosions and airfield damage, and - most alarming of all - a fabricated audio clip in which a senior Pakistani commander appeared to declare a nuclear alert all circulated widely, spreading panic across the region. A deepfake of Pakistan's prime minister, apparently conceding defeat, went viral before it could be debunked. Indian television channels, meanwhile, aired recycled and AI-generated footage as though it were live from the front.

The false claims reached the highest levels. Misinformation circulating during the crisis alleged that India had struck Pakistan's Kirana Hills nuclear site and caused a radioactive leak; a doctored letter to that effect went viral. The Indian defence minister, Rajnath Singh, then declared that the International Atomic Energy Agency should "take charge of the nuclear weapons of Pakistan", and although the IAEA subsequently denied any leak, Singh's intervention led many to believe the incident had genuinely occurred. Elsewhere, an Indian news outlet claimed Pakistan had convened the body that oversees its nuclear arsenal, a claim Pakistan's defence minister had to publicly deny.

This is the nightmare scenario in miniature: synthetic content generating real fear, real fear driving official statements, and official statements lending the synthetic content an unearnt legitimacy. South Asia's deterrence systems rely on short decision windows and tight command-and-control. Compress those windows with a fog of fabricated imagery and a leader may act on a signal that was never sent. The next flashpoint may be triggered by tokens, rather than weapons.

Between Luck and Lies

The disinformation need not even be authored by a human adversary. Increasingly, the machines we have installed as arbiters of fact generate the errors themselves. Following a seismic event in Pakistan last year, users of X turned in large numbers to its built-in chatbot, Grok, to ask whether the tremor might have been an underground nuclear test. Kjølv Egeland, a senior researcher at NORSAR, an independent research foundation that serves as the Norwegian national data centre for the Comprehensive Nuclear-Test-Ban Treaty, tested this out for himself. At first, the results were surprisingly promising: Grok confidently reported to him that the quake had occurred at a depth of ten kilometres - too deep for a nuclear test - and reached the correct conclusion. The trouble was that almost every supporting fact it offered was wrong: it misdated a Soviet test, misstated its depth, and crucially, appears to have lifted the ten-kilometre figure from a placeholder value that seismological observatories use to filter out noise. A subsequent forensic analysis put the true depth at around a hundred kilometres. Grok, Egeland noted, "was lucky this time." Its facts were false; its conclusion merely happened to be right.

It is not difficult to imagine the version where the luck runs out. And the pattern is not confined to one chatbot. When an 8.8-magnitude earthquake off Russia's Kamchatka Peninsula triggered Pacific-wide tsunami warnings in July 2025, chatbots and AI-generated search summaries repeatedly told users that still-active warnings had been lifted; in some cases inventing sources to support the false reassurance. This is unfortunately the very nature of many (if not most, or more probably all) chatbots with which the layperson engages today: prone to hallucinations, seeking to drive the next prompt at the expense of accuracy and, therefore, public safety. A system optimised to please its user is not, at the decisive moment, optimised to tell that user the truth.

The State as Author

Where commercial platforms are careless, states are deliberate. Russia has been the most industrious practitioner. In the autumn of 2022, as documented by the Head of the Odesa Centre for Non-Proliferation Polina Sinovets, Moscow mounted a concerted disinformation campaign alleging that Ukraine was preparing to detonate a "dirty bomb" on its own territory; a fabrication apparently designed to lay the groundwork for Russia's own possible use of a tactical nuclear weapon at a moment when its forces were dangerously exposed. The episode risked tipping into a NATO–Russia confrontation. It was disinformation deployed not to win an argument but to license war crimes and atrocity.

The campaigns have only grown more sophisticated as the tools have improved. At the end of last year, Ukraine's Centre for Countering Disinformation reported that Russian propaganda was circulating a fraudulent report, falsely attributed to the legitimate Institute for the Study of War, claiming that only 30 per cent of France's nuclear forces were operational. The Institute had published no such thing; it doesn't even audit NATO arsenals. The purpose, the Centre explained, was to manufacture an illusion of "NATO weakness", to discredit France and President Macron, and to erode trust in Ukraine's allies. The same watchdog documented AI-generated videos of fictitious "Ukrainian soldiers" pushing narratives of collapsed defences and chaos in the ranks, and earlier false, AI-generated reports designed to convince the public that a contested city had already fallen. A more crude precursor, in March 2022, at the start of the war, had shown Ukrainian President Zelenskyy apparently instructing his soldiers to lay down their arms.

By March of this year, when the United States and Israel attacked Iran, the technique had become a flood. Social networks were overrun within the first weeks of the war by AI-generated videos of missiles striking American warships and explosions convulsing Tel Aviv, a sizeable portion of it Iranian in origin and most memorably in the surreal form of Lego-animated caricatures of President Trump and his cabinet. All of it was optimised for the emotional, spectacle-hungry algorithms of TikTok and X. The fog of war has always existed. What is new is that it can now be manufactured on demand, at industrial volume, by anyone with a smartphone.

Most troubling of all is when the leaders themselves join in. President Trump announced on his own Truth social-media platform in August 2025 that he had ordered two nuclear submarines "to be positioned in the appropriate regions" in response to what he called the "foolish and inflammatory" statements of the former Russian president Dmitry Medvedev. Nuclear submarines cannot (one hopes?) be commanded by social-media post. But a head of state issuing this type of strategic signal on an engagement-optimised platform (and, by implication, consuming his own information within the same clickbait-outrage-driven environment) represents a profound departure from the old expectation that such matters be handled on the basis of verified fact and expert consensus. When the signal and the noise issue from the same channel, the distinction between them begins to dissolve.

The Human in the Contaminated Loop

There is a further, colder danger here: as states fold AI into their defence and intelligence systems, the technology risks becoming not merely a source of confusion but an active accelerant of escalation. A widely reported study from King's College London placed leading chatbots - including systems from OpenAI, Anthropic and Google - in the role of national leaders in simulated war games. The models showed a persistent tendency to escalate, sometimes to use of nuclear weapons, even when de-escalation was available. The researchers suggested the systems treated backing down as reputationally catastrophic and, lacking any visceral horror of nuclear war, reasoned about it in coolly abstract terms.

Herbert Lin, a researcher focusing on the intersection between emerging technologies and US national security, set out the mechanism earlier this year for AI-generated disinformation as "a multiplier of existential threat": AI, he argues, elevates the risk of nuclear escalation by "amplifying disinformation, overloading analysts, compressing decision timelines, and exploiting cognitive and institutional vulnerabilities" in the human-and-machine systems that govern nuclear command and control. It can shape perceptions, contaminate the intelligence feeding a decision, and destabilise the delicate signalling by which nuclear-armed rivals communicate their intentions to one another. Each of these was once the slow work of human spies and propagandists. Now it can be automated by a basic Claude-assisted cron job.

The point is not that machines will seize control from us, but that they will make the humans still nominally in charge faster, more confident and more wrong. We already have a sobering illustration of how thin the margin can be. Shortly after Russia began its "special operation" in Ukraine in 2022, India accidentally fired a BrahMos cruise missile into Pakistani territory; it was only Pakistan's restraint, and a correct reading of the incident as an error rather than an opening salvo, that prevented a response. Now imagine that same accident occurring not in relative information calm but amid a torrent of deepfaked "confirmation" — a fabricated commander's voice, a forged strike order, a synthetic satellite image — each artefact nudging the interpretation towards the worst case, and the window for cool judgement shrinking by the second. The technology does not need to pull the trigger. It only needs to persuade a frightened human that the trigger has already been pulled against them.

A Warning from Rome

If further evidence were needed that this is not the anxiety of cranks, it arrived earlier this month, when the Global Nobel Laureates Assembly on Artificial Intelligence and Nuclear War (rolls off the tongue, doesn't it?) convened in the papal summer villas at Castel Gandolfo, outside Rome. This gathering, largely ignored by mainstream media, brought together more than a dozen Nobel laureates alongside senior scientists, policy experts, and even religious leaders; the first time that so large and varied a body has addressed the twin dangers of AI and nuclear war in concert. On 16th July they signed the Rome Declaration, formally titled Humanity at the Threshold and aligned with Pope Leo XIV's encyclical Magnifica humanitas, warning that the "AI-nuke nexus" was rapidly eating away at human control over existential decisions.

What is striking is how squarely the laureates placed disinformation at the centre of the danger: "AI enables information warfare," the declaration states. "Peace depends on a shared reality. Without facts there can be no truth; without truth there can be no trust; and without trust, no dialogue, no verification, and no restraint among nations." Maria Ressa, a 2021 Nobel Peace laureate and co-author of the text, put it more vividly still: "An atom bomb exploded silently in our information ecosystem. Now, crises can be manufactured in minutes."

That is very nearly the whole of the matter in two sentences. The assembly's recommendations - chief among them halting both arms races and governing AI in line with "humanity's interests, including international law and respect for human rights" - are admirable, if lofty. But the deeper contribution is not so much prescriptive of what should be done, as diagnostic of the overall situation: restraint between nuclear-armed rivals rests on a shared reality, and a shared reality is precisely what these new technologies are undermining.

What Might Be Done

None of this is hopeless, and the same experts who diagnose the problem have begun to sketch responses. Several are worth taking seriously.

The first is to inoculate the public before the fact rather than merely correcting it afterwards. "Prebunking" - exposing audiences in advance to the tell-tale techniques of fabricated damage footage and forged official warnings - has been shown to blunt the spread of the real thing, much as a vaccine primes the immune system. Media-literacy education, sustained rather than sporadic, belongs in the same category.

Secondly is to change the defaults of the platforms themselves. The European Union's Digital Services Act already contains a crisis-response mechanism; extended and applied with transparency, a "crisis mode" could throttle the viral spread of unverified claims in an affected region, elevate authoritative alerts, and suspend the revenue-sharing that currently pays people to farm engagement during a disaster. Provenance standards such as the Coalition for Content Provenance and Authenticity, which embed a verifiable signature recording how an image was made, would help, but only if platforms actually display them, which so far, they largely have not.

Third is to shore up the institutions of verification and communication. Egeland argues for properly funding and amplifying the genuinely authoritative monitors — the IAEA and the network of stations maintained under the Comprehensive Nuclear-Test-Ban Treaty — so that a credible voice exists to be heard above the din. In South Asia, experts are calling for the machinery of de-escalation to be rebuilt and used: revived leader-to-leader hotlines, restored cross-border journalism, bilateral mechanisms for flagging suspected deepfakes, and the quiet Track II dialogues that once fed useful ideas into formal agreements.

And the fourth measure, underlying all the rest, is to keep human judgement at the centre of nuclear decision-making. The consensus emerging from the arms-control community is that nuclear-armed states must retain unambiguous human control over any decision to use these weapons, and should build automatic firebreaks into increasingly automated command systems, precisely so that a compressed timeline and a contaminated feed can never combine to make the choice for us.

The Discourse Is the Battlefield

AI will, in the near future, take an ever-larger role in the targeting, development, and proliferation of nuclear weapons themselves. That is coming, and it deserves the attention it is beginning to receive. But it should not distract us from the threat that is already here, one that requires no breakthrough in machine autonomy, no algorithm handed the launch codes, only the tools that hundreds of millions of us already carry in our pockets.

The first nuclear age peppered an unprepared public with a terrifying new vocabulary and left it, in the old phrase, sadder but wiser. The second age arrives with something the first never had: a machine capable of manufacturing the misleading fog through which every nuclear crisis will now be perceived and, perhaps, misjudged. The discourse around nuclear weapons was always going to matter. It has now become a battlefield in its own right, and we are nowhere near ready for it. Fallout shelters would likely have done little to protect against the bomb, but they are even more powerless to protect against falsehoods.