Every generation produces its own prophets of the end. In the twelfth century it was astrologers reading comets; in 2012 it was a Mesoamerican calendar whose 5,126-year cycle supposedly terminated on December 21st, with crowds gathering at Chichén Itzá and Tikal to watch the world not end. Now it is a panel of global experts, armed with probability estimates instead of hieroglyphs, warning that artificial intelligence and nuclear conflict pose the two greatest extinction-level risks to humanity by 2100. The report deserves to be taken seriously. But it also deserves to be read against the long and humbling history of humans trying to forecast their own annihilation, a history in which the committees are usually right about the danger and wrong about the shape it takes.

The Weapon That Already Exists

Start with what we know rather than what we fear. Nuclear warfare remains, by any sober accounting, the most immediate mechanism by which humanity could engineer its own destruction. A large-scale nuclear exchange would kill hundreds of millions of people directly, and the secondary effects (fallout, nuclear winter, nuclear famine) could push the toll toward civilizational collapse or worse. This is not speculative technology. It is hardware that has existed since 1945, sitting in silos and submarines, requiring no breakthrough, no emergent capability, no committee of experts to debate whether it has arrived. It is already here.

And we have already nearly used it to annihilate ourselves, not once but repeatedly. The October 1962 Cuban Missile Crisis is the canonical case: thirteen days in which the United States and Soviet Union stood closer to nuclear war than at any point before or since. What is less often emphasized is what happened afterward. Both sides, having stared into the abyss, realized how easily a nuclear war could have started and how terrible the results would have been. That shared vertigo produced something genuinely new: the Moscow-Washington hotline, and in July 1963 the Nuclear Test Ban Treaty, under which the USSR, the US, and Britain agreed to confine testing underground. The crisis did not end the danger. But the near-miss itself became the mechanism of partial correction.

This is the pattern worth dwelling on: humanity's survival of the nuclear age has depended less on the absence of catastrophic risk than on a string of near misses, officers who declined to launch, false alarms correctly identified as false, political leaders who blinked at the last possible moment. A "near miss" is defined precisely as an event with the potential to cause catastrophic harm that, for whatever contingent reason, does not occur. The Cold War produced dozens of these, many declassified only decades later. If you build a risk model of nuclear extinction, the historical base rate is not zero near-catastrophes; it is remarkably frequent near-catastrophes, averted by thin margins. A report projecting risk out to 2100 has to grapple with this: the danger is not a tail risk sitting quietly in the distribution. It has already brushed past us several times.

The Problem With Naming the Next Threat

Artificial intelligence is a different kind of risk entirely, and this is where the 2100 report runs into the oldest epistemic trap in the business of doomsaying: the assumption that today's most alarming technology is necessarily tomorrow's instrument of extinction. Every era's experts have looked at their own era's frontier technology and concluded it was the one that would finally do us in. In the 1970s and 80s it was overpopulation and resource exhaustion that dominated extinction literature. By the 2000s, climate change had taken that seat; Sir John Houghton, former head of the UK Meteorological Office, described it explicitly as a weapon of mass destruction, one that "knows no boundaries" and "can strike anywhere, in any form, a heatwave in one place, a drought or a flood or a storm surge in another." Gallup polls across 127 countries in 2007 and 2008 found that more than a third of respondents already regarded it as an existential-scale threat to their way of life.

None of this means climate change or AI risk is overstated. It means the specific mechanism experts fixate on tends to reflect the anxieties of their moment rather than a stable forecast of what actually threatens the species. The 2012 phenomenon is almost a parody of this tendency: an entire global subculture convinced that an ancient calendrical cycle, rather than any of the genuinely dangerous systems humanity had built, would be the agent of apocalypse. It wasn't. The calendar simply rolled over, as calendars do. The lesson is not that doomsday talk is always foolish; nuclear weapons were, and remain, a real and underappreciated danger throughout the Cold War precisely because too many people treated deterrence as a stable equilibrium rather than a probabilistic bet renewed daily. The lesson is that specificity about mechanism is usually where these forecasts go wrong, even when the underlying sense of peril is justified.

Why Committees Struggle With Compounding Risk

There is also a structural reason to be skeptical of any single report's ability to rank-order existential threats. Expert panels are good at identifying categories of danger and reasonably good at estimating single-variable probabilities. They are far worse at modeling how risks compound: how an AI-accelerated disinformation campaign might degrade the command-and-control systems that have, since 1963, kept nuclear near misses from becoming nuclear wars; how a climate-driven famine, intensifying the droughts that have already devastated parts of Africa and fed the cycle of malnutrition and disease, might destabilize a nuclear-armed state's government at precisely the wrong moment; how the negotiated-peace option that existed in Libya in March 2011, when the International Crisis Group proposed a ceasefire framework that was never seriously attempted, illustrates how often humanity's crises are not acts of impersonal fate but failures of political will to choose the harder, slower path over the dramatic one.

That last point matters more than it might seem. The Libyan case is instructive precisely because it was not a technological risk at all. It was a human choice, made under pressure, to pursue a path that foreclosed negotiation in favor of intervention, with consequences that rippled through the region for a decade. Extinction-risk reports, by focusing on technologies and AI and nuclear weapons, implicitly frame the danger as something humanity suffers rather than something humanity chooses and re-chooses at every decision point. The hotline installed after 1962 worked not because it was clever engineering but because two governments decided, repeatedly, afterward, to use it rather than ignore it.

What the Report Gets Right

None of this is an argument for complacency. The honest case for taking AI and nuclear risk seriously by 2100 is simply that both are now simultaneously live. Nuclear arsenals have never been dismantled to the degree the Test Ban Treaty's signatories once hoped; they have in some regions expanded and modernized. And artificial intelligence, unlike the comet-watchers' prophecies or the Mayan calendar, is a technology actually accelerating in capability in real time, with genuine uncertainty about how it interacts with existing command structures, information ecosystems, and geopolitical tensions. The danger is real. What should be resisted is the false precision of naming 2100 and these two mechanisms as though the committee has solved a forecasting problem that no generation before it has managed.

The Better Question

The historical record suggests a different emphasis than the one these reports usually take. Instead of asking which technology will end us, the more useful question is whether we are building the equivalent of the Moscow-Washington hotline for each new risk as it emerges: institutional mechanisms for de-escalation, verification, and course correction that do not depend on getting every individual judgment right, only on surviving the near misses that are now certain to come. The Cuban Missile Crisis did not end the nuclear age. It produced forty years of continued risk, managed imperfectly, through institutions built in its aftermath. If AI poses a comparable risk, the honest response is not a 2100 extinction date but an equivalent, urgent effort to build the guardrails before the first serious near miss, rather than, as happened with nuclear weapons, only after it.

Humanity has a long and uneven record of surviving its own inventions, not because it correctly predicted which invention would be fatal, but because it built mechanisms, treaties, hotlines, and test bans to absorb the near misses it did not predict. That is a thinner reed of comfort than a committee report would like to offer. It may also be the only honest one.