A half-joking interface for a serious uncertainty · v1.2.0
"Apocalypse" is not an event but a scale. Three thresholds, each with its own criterion, its own probability and its own dominant mechanism: human malice rules the bottom of the ladder, control failure the top. The model computes not a date but a risk intensity, and from it a cumulative probability and a median date.
ASSUMED. An expert judgement by the author of the model. Contestable by design. How readily people turn available capability to harm. A model input on an arbitrary scale, not a share of users, models or requests. Main driver of the bottom rung.
ASSUMED. An expert judgement by the author of the model. Contestable by design. How strongly loss-of-control scenarios feed the risk. A model input, not a probability that AI loses control: it is multiplied by rung weight, capability, wiring and mitigation before it means anything. Main driver of the top rung.
ASSUMED. An expert judgement by the author of the model. Contestable by design. Regulation, audit, kill switches, treaties. Not a percentage of risk removed: every rung has its own ceiling, and the implied reduction is shown below.
ASSUMED. An expert judgement by the author of the model. Contestable by design. What share of energy, finance, logistics, weapons and medicine is already handed to autonomous loops.
ASSUMED. An expert judgement by the author of the model. Contestable by design. How many years it takes for wiring to reach its ceiling.
ASSUMED. An expert judgement by the author of the model. Contestable by design. Once capability plateaus, the world learns to live with it and risk decays over this period. Set it to 100 and it never decays.
Each rung is the probability of an event at that level or worse, so the bottom rung always sits above the top one. 34% local includes the 3% global, it does not replace it.
≥ 1,000 dead or ≥ $10bn in damage. One city, one network, one company.
Oct 2041 median
P by 2035 29.8% · 2050 67.2% · 2100 88.1%
deaths 1K–100K · damage $10.0B–$100.0B
Failure of an autonomous loop in regional infrastructure; a model-assisted attack on an organisation’s network; a laboratory incident.
≥ 1 million dead or ≥ $1tn. National infrastructure failure that takes years to undo.
> 2100 P < 50% within the horizon
P by 2035 9.5% · 2050 29.8% · 2100 49.9%
deaths 1M–100M · damage $1.0T–$10T
Cascading failure of financial and energy systems; conflict escalation with AI inside the decision loop; biological risk amplified by access to design tools.
≥ 10% of the world population, or humanity irreversibly losing control of its own future.
> 2100 P < 50% within the horizon
P by 2035 2.7% · 2050 10.6% · 2100 20.1%
deaths 800M–8B · damage $100T–$1000T
Control failure in a system wired into every critical loop at once; irreversible concentration of power; a scenario nobody described in advance — which is what makes it global.
Cumulative probability of an event at each level
P(an event at this level or worse has occurred by year t), counting from today. Each curve includes everything above it on the ladder, so the bottom rung always sits above the top one. The plateau on the right follows from the window of vulnerability: survive the transition and risk falls.
Cumulative probability by 2050 and 2100: Local: 67.2% by 2050, 88.1% by 2100. Regional: 29.8% by 2050, 49.9% by 2100. Global: 10.6% by 2050, 20.1% by 2100.
The chart is keyboard focusable: arrows move the cursor, Shift jumps ten years, Esc clears it.
Expected toll by 2050
269.7M
expected deaths: summed over three levels, probability × geometric mean of the range
$34T
expected direct damage in 2026 dollars
This is arithmetic, not a forecast, and it is not a claim that anybody will die. It is the probability of an event at each level × the geometric mean of the casualty band that level is defined with, summed over the three levels — so it inherits every assumption on this page and adds a rounding of its own. A cumulative probability is not a count of events, and each rung is defined with an OR, so an event can clear the damage threshold while killing nobody. Read it as an order of magnitude produced by the sliders you set, and read the ranges below for what the model actually distinguishes.
Illustrative severity by 2050
| Rung | P by 2050 | deaths | damage |
|---|---|---|---|
| Local | 37.4% | 1K–100K | $10.0B–$100.0B |
| Regional | 19.2% | 1M–100M | $1.0T–$10T |
| Global | 10.6% | 800M–8B | $100T–$1000T |
Ranges, not an expected value. A single number would hide two things: the cumulative probability of an event is not a count of events, and every rung is defined with an OR — an event can clear the damage limb while killing nobody.
Each row is the probability of an event at exactly this level and nothing worse, which is why the rows add up to the probability of an event at any level. It is the difference between two rungs of the cumulative table above: P(this level or worse) − P(the next level or worse). So "exactly local" is lower than "local or worse": the worlds where a regional or global event also happened are counted in those rows, not in this one.
Implied risk reduction
The slider is a strength, not a percentage of risk removed. Each rung caps how much mitigation can achieve, and the cap is lowest where the event has never happened before.
λᵢ(t) = (malice·wᵢ + control failure·uᵢ) · cᵢ(t) · d(t) · (1 − mitigation·eᵢ) · aᵢ(t)
cᵢ(t) is capability: a logistic function of the log horizon relative to the rung threshold. d(t) is wiring, a saturating exponential. eᵢ is the mitigation ceiling: 0.80 local, 0.60 regional, 0.45 global. aᵢ(t) is decay after capability saturation. Result: P = 1 − exp(−∫λ dt), integrated in yearly steps.
Note that a product of four multipliers, each of which you set by eye, yields a quantity accurate to an order of magnitude at best. Every decimal place in these dates is a polite lie told by the interface.