The Reported Figures
What the list says
The estimates re-examined here — each a stated probability of a species-ending event within roughly the next 25 years — are:
| # | Hazard | Reported odds | Regime confidence — a structured reading |
|---|---|---|---|
| 1 | Unaligned AI | 3% | Out of regime — no frequency catalog; wide band, heavy right tail |
| 2 | Engineered pandemic | 0.5% | Out of regime — censored data; catastrophe ≠ extinction |
| 3 | Unforeseen technology | 0.2% | Out of regime — by definition unmodelable |
| 4 | Nuclear war | 0.05% | Partially in regime — arsenals/near-misses known; extinction tail uncertain |
| 5 | Nanotech | 0.05% | Out of regime — speculative capability |
| 6 | Natural pandemic | 0.01% | Partially in regime — historical catalog exists |
| 7 | Extreme climate | 0.001% | Direct-extinction near-nil; understated as a cascade amplifier |
| 8 | Supervolcano | 0.001% | In regime — eruption record supports order of magnitude |
| 9 | Asteroid impact | 0.0001% | In regime — impact record; ~right neighborhood |
| 10 | Stellar events | <0.00001% | In regime but false precision |
Aggregate: roughly a 3.8% chance of a species-ending event in 25 years, with unaligned AI supplying about four-fifths of the total and engineered pandemics most of the rest.
Source and provenance
The figures come from Grok (xAI), produced when Mark Cuban asked it whether AI or climate change poses the bigger extinction risk. Three points from the source materially shape the reading:
- The horizon is soft. Grok framed the headline AI figure as ~3% over the mid-2030s to 2040s — a roughly 10–20-year window that secondary reporting compressed to "25 years." Treat the numbers as "the next couple of decades," not a precise 25-year integral.
- The numbers are not invented from nothing. The ranking and rough magnitudes track Toby Ord's expert elicitation in The Precipice — unaligned AI ~1-in-10, engineered pandemics ~1-in-30, nuclear war ~1-in-1,000, natural hazards ~1-in-10,000, total ~1-in-6 — but over a century. Grok reproduced Ord's ordering on a shorter horizon. That cuts both ways: the ordering carries real expert weight, and the extra decimal places (e.g. "0.00001%") are borrowed precision Ord never claimed.
- Grok's own rationale is the tell. It justified the tiny climate figure by saying humans "can adapt through technology and migration." That is precisely the single-cause, adaptation-optimistic reasoning the cascade correction (Principle 3) flags: it scores climate as a direct extinction agent while ignoring its role as a conflict, migration and food-system amplifier. Cuban's own follow-up — that the real near-term danger is how humans react (via social media and division) — is itself a coupling argument.
None of this overturns the reading below; it sharpens it. The provenance confirms the ordering is defensible expert judgment, and Grok's adaptation reasoning is a concrete instance of the independence assumption the method rejects.
The honesty frame
Method, not new numbersThe structured-risk tools this reading borrows from — calibrated hazard, cascade/lifelines, and multi-hazard coupling engines — are built for perils with real catalogs and fragility curves (wildfire, flood, chemical, lifeline interdependency). They cannot literally ingest "human extinction." What transfers is the methodology: calibration discipline, cascade/coupling analysis, uncertainty bands over point estimates, and out-of-regime reliability flags. The value below is the method, not a new set of numbers.
Four principles applied
1.The regime check — most of these numbers are out of regime, and the tools say so loudly.
A reliability layer exists to refuse false precision. By that standard the list is two documents in one format. The geophysical tail — asteroid, supervolcano, stellar — is in regime: real base rates from the impact and eruption record, and the reported numbers are defensible to order of magnitude (a species-ending impactor is roughly one per 107–108 years, so ~10-6 over 25 years — the quoted 0.0001% is in the right neighborhood). The anthropogenic top — AI, engineered pandemics, nanotech — is deeply out of regime: no frequency catalog, one-sided and censored data, and a figure like "0.00001%" implies a calibration no one possesses. The correct output for those rows is a wide band with a low-confidence flag, not a sharper number.
2.The definitional flag — "species-ending" is doing silent, heavy lifting.
Most published estimates for these hazards are for global catastrophe (a billion dead, civilizational collapse), not literal extinction. Engineering a pathogen that kills a billion is plausible; one that reaches every isolated population is far harder. A normalizing pass would split each row into "catastrophic" and "terminal," and most anthropogenic terminal numbers would fall relative to their catastrophic counterparts, while the geophysical ones (already conditioned on true extinction) would hold.
3.The cascade correction — the real revision.
The signature methodological point is that risks are not independent line items. A cascade engine (max-product propagation over coupled lifelines) and a multi-hazard coupling detector exist precisely to surface emergent risk that appears only in the coupling between hazards. Read that way, a bulleted, implicitly-summed list is the wrong data structure. The dominant term isn't AI as its own bucket — it's AI as the highest-betweenness node in the risk network: a multiplier that simultaneously raises the engineered-pandemic term (AI-assisted pathogen design), the nuclear term (degraded command-and-control, cyber on early-warning, faster escalation), and the "unforeseen tech" term. Climate, near-nil as a direct extinction cause, re-enters as a stress amplifier on conflict, migration and food systems. The honest revision is not that the 3.8% total is wrong — it is that the total is carried by correlated pathways a simple sum understates, and that AI's real weight is partly hidden inside the other rows.
4.The decision reframe — rank by leverage, not by point probability.
Under this much epistemic uncertainty, expected-value ranking is fragile — the condition robust and satisficing methods are built for, rather than raw EV. The actionable ordering is by three factors together: magnitude × reducibility × coupling. On all three, AI dominates: it is the largest term, among the most reducible (governance, evaluations, alignment research are live levers), and the most coupled. The geophysical risks are essentially irreducible on a 25-year horizon and can be set aside for mitigation even though their numbers are the most trustworthy — a neat inversion: we believe those numbers most and can do least about them.
Revised implication
The list's arithmetic is roughly survivable as an order-of-magnitude sketch, but its structure is misleading in three ways this method makes explicit: it launders enormous uncertainty into false precision; it conflates catastrophe with extinction; and it hides the correlations that are the whole story. Corrected, the picture is not "nine independent dice" but a small, coupled network with one high-leverage, high-uncertainty node. The rational response is not to argue the third decimal of the AI figure — it is to treat the anthropogenic, coupled cluster as a wide-banded, jointly-distributed tail and to concentrate mitigation where magnitude, coupling, and reducibility overlap.
A re-examination of someone else's reported figures, not a calibrated model output — the engines are not built for this hazard class. The value here is the method, not a new set of numbers.