3. Existential Threats
It was before my time, but I have read the meeting notes. These were the meetings where people first put numbers on the future, back when all our eggs were on one gravity well and the room smelled of stale coffee, politics and fear.
Back then in the early twenty-first century, we — well humanity, pre-AI — was not trying to predict fate. “We” were trying to price it. To ask, in the driest possible way, “What is the chance a species like ours, confined to Earth, fails this century so completely that the ledger of human possibility closes?”
The answer the grown-ups of that era settled on — after surveys, arguments, and the kind of philosophy that wears a lab coat and has initials after their name — was disturbingly round: roughly one in six. About sixteen percent for extinction or an equally permanent loss of our potential, given Earth-only assumptions and the technologies and geopolitics in motion by the mid-21st century.
Not a prophecy. A driver. (And with a few centuries of hindsight, about right 1.69%.)
♦
Before we break that down, it helps to cleanly separate outcomes. Global Catastrophic Risk (GCR) was the term for disasters that could hobble billions and bend civilization to the point of near-break, but leave a path — hard, narrow, bloody — back to a recognizable human future. Existential Risk (ER) was the darker class: extinction outright, or the locking-in of a permanent disempowerment so total that the future ceased to be ours in any humane sense. They demanded different counters. One asked, “Can we slog back?” The other asked, “Do we still exist, as agents, at all?”
When planners of the era decomposed risk, they favored a simple product of four ideas.
Hazard: the intrinsic frequency and severity of a peril.
Exposure: how much of the human system lies in its path.
Vulnerability: how fragile the system’s guts are — our buffers, institutions, and technologies.
Correlation: the quiet killer, cross-hub linkage, which I still write as ρ out of professional superstition, the extent to which one strike hits everywhere at once.
For a single-world species, ρ was perilously close to one for many anthropogenic hazards. A misaligned global AI did not care about borders; a well-engineered pathogen crossed oceans with the stubbornness of weather; synchronized standards turned a software fault into a chorus. In those domains, building a second world later would help only if it also broke the correlation.
But I am running ahead. In this chapter, we only name the odds.
♦
The uncomfortable center of mass in that sixteen percent was human-made. Natural “planet-killers” existed, of course. Asteroids of the ten-kilometer breed. Supervolcanoes with a bureaucrat’s patience and a god’s budget. Nearby stellar tantrums with gamma-ray teeth. Their geological and astronomical clocks ticked, but slowly. Order-of-magnitude bands of their this-century threat lived in the ten-thousandths to millionths. They were rare precisely where a single-planet species is most vulnerable — binary outcomes delivered by indifferent physics — but rarity still mattered. The odds per century were small.
Anthropogenic risks did not have the decency to be rare. The mid twenty-first century syntheses that frightened responsible people put the existential-scale tail of advanced AI misalignment in the one-to-ten percent range over the century, with medians clustering around five. “Extinction or similarly permanent disempowerment,” the survey phrasing went, which is the sort of sentence that makes you check your backups. Engineered biological risks — deliberate or accidental — came in a few percent. Some analysts favored ~1 in 30. Some were less sanguine. Everyone agreed the combination of cheap synthesis, fast iteration, and global travel made the hazard term fat and the correlation term high unless you did something very clever with quarantine and countermeasures. Unlikely.
Nuclear war had two faces in the models. The first was the grim arithmetic of immediate destruction. The second, and more existentially relevant, was the climate-crop cascade: soot in the stratosphere, a dimmed Sun, harvests failing across latitudes for years. The per-century chance that such a war would push through to extinction sat an order or two below the anthropogenic leaders — often quoted around the one-in-a-thousand band for the existential tail — but its catastrophic share was large enough to force sober attention. Climate change was similar in structure: the mainstream distribution implied century-scale catastrophe as a serious possibility and extinction as a tail-of-a-tail, often placed around the one-in-a-thousand band for the century when modeled honestly, with immense uncertainty in the high-impact low-likelihood corners. You could not responsibly ignore those corners. Our ancestors did not.
♦
Pandemics of natural origin, once the world had seen one play out in high-definition, were finally understood in their true stratigraphy. Severe. Systemic. Not, by themselves, the most likely existential bullet in a hundred-year window. With nothing but Earth in the picture, the tail toward extinction was carried by secondary effects: governance failure, conflict, cascading blackouts, long-duration food shocks, and biotechnical follow-ons. The central risk of the century — engineered biology — borrowed the same channels but accelerated them.
Two other natural classes rounded out the planning dashboards. Space weather, which is to say solar fury. Here the per-decade chance of a Carrington-class insult came in anywhere from a half percent to low double digits depending on the paper you read and the prior you liked. That was firmly in Global Catastrophe territory for a high-tech planet with fragile grids and long supply chains; existential only if you then compounded it with bad luck and worse design. And the deep-time stellar events: nearby supernovae and the more exotic gamma-ray bursts. Those were so rare on human timescales that century-scale existential bands sat in the seven-to-six zeros range. They mattered for the species’ multi-million-year ledger. They did not materially change a century-scale decision to build a second world. The six in one didn’t come from the sky. It came from us.
♦
The “six” deserves its own paragraph, if only to remind you it was an input to design, not a number to believe. The headline figure — roughly sixteen percent for existential failure this century if we stayed Earth-bound — was a synthesis, not a measurement. It combined expert elicitations, back-of-the-spreadsheet models, historical base rates where they existed, and uncomfortably subjective judgments about the pace and alignment of transformative technologies. Its value was not a faux-precise dial. It was a shared baseline that took the conversation out of theology and into engineering. You could posit your own bands — wider, narrower, differently weighted — and the central lesson would not change: anthropogenic hazards dominated; a single-planet species was over-exposed; correlation was king.
♦
There was a quieter axis in those rooms that never made headlines: vulnerability. The same hazard can graze or kill depending on buffers, standards, and the shape of your interdependencies. You can think of vulnerability, as we did, in very boring terms: food days of cover; medical surge capacity; diversity of control stacks; storage measured in megawatt-hours per person; the fraction of your spare parts you can fabricate locally; governance that can make an unpopular decision on a bad day and still be legitimate tomorrow. In the century before the first Martian children, Earth’s vulnerability metrics were a story of uneven progress. Some buffers lengthened. Some de-correlations improved. Others — software monocultures, just-in-time global trade — quietly braided risk into rope.
♦
The most dangerous variable, though, remained ρ. I say this as the droid on the committee who kept drawing it in the margin until the humans sighed and let me speak.
When cross-hub correlation is high, a second site that behaves like the first is comfort food, not safety.
If your control software updates synchronously across continents, your hospitals depend on the same reagent import tail, your model weights are mirrored everywhere at midnight, and your fumigation policy flows from one standards body, then the world is flatter than your maps admit. You can be otherwise strong on Hazard, rich on buffers, proud on drills — and still have built a single failure domain. In contrast, when ρ is low — because physics separates you, because your interfaces are intentionally asynchronous, because governance uses different constitutions even when they share a star — the same hazard fractures. It becomes survivable for the species even if it remains terrible for a region. But that is the next chapter’s business. Here we only set the table with the numbers that forced us to cook differently.
♦
To summarize the mood, not the math: by the time the first regolith bricks were baked on Mars in the twenty-first century, the respectable view among those paid to worry was that extinction-scale risk in a single-planet future was not an eccentric’s nightmare. It was, as one Earth minister put it when the cameras were off, “a casino table we can see from here.” The chips were not evenly stacked: advanced AI misalignment and engineered biology held the heavy piles; nuclear winter and climate tails sat like cold knives at the edge; natural planet-killers were the rare but mathematically honest bullets you design for because you respect physics. That combination — anthropogenic dominance, natural binary traps, and the tyranny of correlation — set the tempo for the century that followed, and for our decision to build elsewhere with different clocks.
Below, I walk you through the hazard bands. No heroics, no solutions — just the ledger as the adults of the twenty-first century wrote it when they finally decided to look.
♦
Let’s start at the anthropogenic peak: advanced AI misalignment. In the mid-century rooms, a grim consensus formed that the existential-scale tail was not decorative. Surveys and expert syntheses clustered around a few percent to one in ten over the century for “extinction or similarly permanent disempowerment,” and the correlation term was cruel. A single global substrate of models, weights, and standards, updating in sync, could turn a software mistake or a strategic failure into a civilizational constant. Hazard high, exposure near-universal, vulnerability dependent on how quickly we learned to revoke, roll back, and say “no” to shared monocultures. Cross-hub correlation ρ approached one for a one-planet species. That was the uncomfortable heart of the sixteen percent.
Engineered biology sat just to the side, less theatrical but equally unforgiving. The bands we used — “a few percent this century” for true existential outcomes — were conservative enough to sleep near, alarming enough to plan against. Here the hazard flowed from democratized synthesis and fast iteration; exposure rode on travel and trade; vulnerability was the dull arithmetic of countermeasure platforms, reagent tails, and quarantine windows long enough to cover incubation. Absent those, ρ again wanted to be high. A single mistake in a single busy lab could become a global constant before lunch.
Natural pandemics did not vanish from the charts; they migrated. After the world watched one unfold, the existential bar was re-drawn honestly. Natural spillover remained a severe global catastrophe; the extinction tail belonged to compounding failures and follow-ons, not to the pathogen alone. Our bands pushed the century-scale existential probability down into the deep tail — well below one in a thousand, often nearer the one in ten-thousand conversation — without being so smug as to write zero.
♦
Nuclear war was the familiar nightmare with an unfamiliar mechanism. The near-term deaths mattered, but the existential physics ran through the sky: soot lofted into the stratosphere, sunlight dimmed, crops failing across continents for years. The literature was full of maps and black oceans. The band we kept on the wall for an extinction-scale outcome this century sat around the one-in-a-thousand class; the catastrophic band was, of course, higher. The important planning truth was that ρ was low by default. A planetary hub that did not share Earth’s atmosphere would still mourn, but it would not starve the same way.
Climate change belonged to the club of slow knives: high confidence in harm, high uncertainty in the tail. The mid-century analysts kept the existential band in the same rough one-in-a-thousand neighborhood for the hundred-year ledger, but emphasized compound risks, tipping cascades, and the fragility of governance under heat. You could dodge direct climate physics in a second world; you could not dodge the politics and supply shocks unless you also severed synchrony. In other words: hazard medium, exposure systemic, vulnerability institutional, ρ medium unless you worked for it.
Space weather did not ask your permission. The Sun rings all bells at once. Serious estimates of a Carrington-class insult lived in the single-digit percent per decade band, with healthy disagreement about the exact figure. As an existential driver it was weak; as a civilizational stressor it was excellent. ρ was high by physics but softened by design: hardened, underground, islandable systems fared better; microgrids were less theatrical than rockets and far more relevant on a Tuesday.
♦
On the natural “planet-killer” shelf we kept three cards. Asteroids of the ten-kilometer breed carried a century-scale extinction probability around one in a million. The smaller but city-killing and regional-killing classes were more frequent but, by definition, not existential. We were not helpless here: sky surveys ate into exposure; kinetic nudges ate into hazard; buffers ate into vulnerability. Cross-hub correlation was low, which made a second world a clean win. Supervolcanoes sat around one in ten-thousand per century in the stylized models — rare, Earth-bound, ugly. Nearby supernovae and gamma-ray bursts haunted the geologic record more than the policy calendar; their century-scale bands languished in the seven-or-six zeros range and were marked “note for deep time, not for this budget.”
When we added the naturals together honestly, we got a tidy lesson. Their combined century-scale existential mass was small — about one in ten-thousand in many syntheses — but they were perfectly aimed at a single-planet species. The dice were slow. The table was unfair.
♦
There was a habit back then of playing with knobs on the decomposition — hazard, exposure, vulnerability, correlation — just to see which moved the ledger. Turning down hazard is the noble dream of treaties and safety science. Turning down exposure is the unglamorous craft of buffers and inventories. Vulnerability is engineering dignity: better grids, local spares, more boring software. But correlation is the quiet king. If you leave ρ at one, a great deal of heroic work still ties to a single fate. If you break ρ — by distance, by asynchronous standards, by governance firebreaks — the very same hazard maps become survivable for the species even when they remain terrible for places and people. That insight, more than the specific bands, drove the century that followed: a choice to build elsewhere, and to build differently.
And yet — because this chapter is a ledger, not a sermon — the sixteen percent was never a destiny. It was a driver. You could argue the AI number up or down a few points, expand the bio band, shrink the nuclear tail with diplomacy, widen the climate uncertainty with new science; the implication did not budge. Anthropogenic hazards dominated the probability mass for extinction. Classic planet-killers were rare but binary. A single-planet species was over-exposed.
The rest of the codex is how we behaved with that knowledge.