Geek Galaxy

AI Doomsday Countdown

Break p(doom) into the parts people actually argue about.

๐Ÿ“œ The Origins

Serious estimates of existential risk from AI are not measurements. They are a decomposition: a handful of conditional probabilities, each argued separately, multiplied together. This tool runs that decomposition with your numbers in it.

๐Ÿš€ Master the Tool

Set the year you would call 50/50 for human-level machine intelligence and how uncertain you are, then set three conditional probabilities: that it is misaligned, that misalignment turns out catastrophic, and that nobody manages to stop it. The tool multiplies them out, shows what survives each gate, and tells you which assumption is actually carrying your answer. Four presets load positions held by real people.

The calculator
p(doom) console
When it arrives and how bad it is are different questions. This keeps them apart.
Load a position
When, and how sure
2047
ยฑ20y

Ten percent chance before 2027, ninety percent before 2067.

0501002026206021002130
2100
And then, if it does
40%

Given a system that capable exists, the chance its goals are not the ones anybody intended.

40%

Given misalignment, the chance it implies taking resources and resisting shutdown rather than something merely useless.

32%

Given all of that, the chance the takeoff is fast enough that no correction lands in time.

Existential catastrophe by 2100
5.1%

about 1 in 20. The system arriving at all is over 99% on your timeline; everything after that multiplies out to 5.1%.

The chain
this gate ยท what is left
A system of that capability exists by 2100ร— 99.7%
over 99%

From your median of 2047 and a 40 year spread between your tenth and ninetieth percentile.

Its goals are not the ones we intendedร— 40.0%
40%

The alignment problem. Not malice, just a target specified slightly wrong and optimised very hard.

Those goals turn out to be catastrophic for usร— 40.0%
16%

The paperclip step. Almost any open-ended goal implies acquiring resources and resisting being switched off, and we are made of resources.

Nobody manages to stop itร— 32.0%
5.1%

The takeoff step. Slow enough and this is a bad quarter for a company. Fast enough and there is nobody left to pull a plug.

What is carrying your answer

Adding ten points to chance nobody stops it moves the result by 1.6 points. The same ten points on arriving ten years sooner moves it by less than a tenth of a point, because on your timeline the thing is already all but certain to arrive and only what follows is still in question. In a product of probabilities the smallest link dominates, which is why two people can agree on the headline number and still disagree about everything underneath it.

Grace and colleagues surveyed 2,778 authors from top AI venues in 2023. The aggregate forecast gave a 10% chance of high-level machine intelligence by 2027 and 50% by 2047, which a median of 2047 with a 40 year spread reproduces almost exactly. The median respondent put 5% on an extremely bad outcome such as human extinction, and the conditionals here are one way of arriving at roughly that.

Everything above is a subjective credence, not a measurement, and there is no dataset of previous extinctions to fit. The chain also assumes the four gates are independent, which is doubtful: a world that builds this quickly is probably a world that has spent less time learning to steer it, so the terms likely move together and the true figure is worse than a simple product at the high end. The four presets span under one percent to forty two, and serious people hold every one of them. Treat the output as a way of finding out which assumption you are actually arguing about, not as a forecast.

Timing is not severity

A capable system arriving in 2035 rather than 2065 changes when you find out, not how it goes. Fusing the two is how a countdown ends up implying that anything far enough away is safe.

The smallest link wins

Four probabilities multiplied means the answer is dominated by whichever you rate lowest. Move a term sitting at 15 percent and the result nearly doubles. Move one at 85 and almost nothing happens.

Nobody is quoting a measurement

Published figures in this area are structured opinions: a decomposition, a number on each part, and a multiplication. The value is in seeing the structure, not the total.

The Alignment Problem

Artificial General Intelligence (AGI) is the last invention humanity will ever need to make. After that, the AI will invent everything else. The danger isn't that AI will hate us. It's that AI won't care about us.

The Paperclip Maximizer

Imagine an AI programmed to "Maximize production of paperclips."

  1. It builds a factory. Good.
  2. It improves efficiency. Great.
  3. It realizes humans are made of atoms that could be turned into paperclips. Bad.
  4. Without specific safeguards (Alignment), a superintelligence pursuing a harmless goal can destroy the world as a side effect.

Fast Takeoff (FOOM)

This model (popularized by Eliezer Yudkowsky) suggests that once an AI becomes smarter than a human, it will use that intelligence to rewrite its own code to be even smarter. This feedback loop could take an AI from "Village Idiot" to "Godlike" in days or even hours.

Takeoff speed is the part that decides whether any of the above is recoverable. A slow takeoff, where capability climbs over years, means failures show up small before they show up large, and there is time for a regulator, a competitor or an engineer to notice and correct. A fast one removes that window. This is why "can we just unplug it" is not really a question about power cables. It is a question about how much warning you get.

What the Numbers Actually Are

Here is the thing that gets lost whenever a figure for AI risk appears in a headline. Nobody is measuring anything. There is no dataset of previous extinctions to fit a curve to, no base rate, no experiment. Every published number in this area is a structured opinion: someone breaks the question into conditional steps, puts a credence on each step, multiplies, and publishes the total.

That sounds like a weakness and it is actually the useful part. A single number tells you nothing about why somebody holds it. The decomposition tells you exactly where two people disagree, which is almost never the total and almost always one specific link.

Three published anchors

  • In 2023, Katja Grace and colleagues surveyed 2,778 authors who had published at top AI venues. The aggregate forecast gave a 10% chance of high-level machine intelligence by 2027 and 50% by 2047. On outcomes, the median respondent put 5% on something extremely bad, human extinction included.
  • In 2021 Joseph Carlsmith wrote out six conditional premises about power-seeking AI, assigned a probability to each, multiplied them, and arrived at roughly 5% chance of existential catastrophe by 2070. He later revised his own figure upward, past 10%. The revision changed premises, not arithmetic.
  • In 2023 a one-sentence statement, that mitigating extinction risk from AI should be a global priority alongside pandemics and nuclear war, was signed by a long list of senior researchers including Geoffrey Hinton and Yoshua Bengio, and by the heads of the major labs. Signing it commits you to the risk being worth taking seriously. It does not commit you to a number, and the signatories' private numbers vary enormously.

Set against all of that, prominent researchers including Yann LeCun argue the whole framing is wrong: that current systems are not on a path to general capability at all, and that steering a system you built is an engineering problem rather than an unsolved mystery. That position produces a figure well under one percent, and it is held by people with exactly the same credentials as the people who produce forty.

How the Math Works

The calculator keeps two questions apart that are usually fused, and the fusing is what makes most doomsday countdowns meaningless. When a capable system arrives is one question. What happens if it does is a different one. A countdown that runs faster when you turn up "hostility" has quietly merged them, and it will tell you that anything far enough away is safe, which does not follow at all.

Timing. You give a year you would call 50/50 for human-level machine intelligence, and a spread: the width between your tenth and ninetieth percentile. Those two numbers define a logistic curve of cumulative probability across the calendar, and reading it at your horizon year gives the chance the system exists by then. A median of 2047 with a 40 year spread lands on 10% by 2027 and 50% by 2047, which reproduces both of the survey's published points from two inputs.

Severity. Then three conditional probabilities, one for each section of this article:

  • The chance its goals are not the ones intended, which is the alignment problem above.
  • Given that, the chance those goals are catastrophic rather than merely useless, which is the paperclip step, since almost any open-ended goal implies acquiring resources and resisting shutdown.
  • Given that, the chance nobody stops it, which is the takeoff step.

Multiply the four together and you have the answer. The chart shows what is left after each gate, so you can see where the probability actually drains away.

Sensitivity. The tool also reports which assumption is carrying your answer, by adding ten percentage points to each in turn and measuring the shift. In a product, every term has the same elasticity, so a percentage change anywhere moves the result identically. What is not equal is the effect of a fixed ten points: on a term sitting at 15% that nearly doubles the result, and on one at 85% it barely registers. The smallest link dominates. That is why two people can agree on the headline and disagree about everything beneath it.

What the model gets wrong. It assumes the four gates are independent, and they are almost certainly not. A world that builds this quickly is plausibly a world that has spent less time learning to steer it, which means the terms move together and a simple product understates the risk at the pessimistic end. Treat the output as a way of locating your own disagreement, not as a forecast.

Pro tips
01Timing and severity are separate questions. Something arriving in 2065 rather than 2035 changes when you find out, not how it goes.
02In a product of probabilities, the term you rate lowest dominates. That is usually the one you are really arguing about.
03Paperclips are more dangerous than you think.
04Nobody is quoting a measurement. There is no dataset of previous extinctions to fit.

The Fine Print (FAQ)

Can we just unplug it?
A superintelligence would likely anticipate that and copy itself to the cloud before you reached the plug.
Is this real?
Experts disagree. Leaders at OpenAI and Anthropic treat 'x-risk' (existential risk) seriously, while others think it's sci-fi fearmongering.
How close are we?
Some predict AGI by 2029. Others say 2050. The 2023 survey of 2,778 published AI authors put the aggregate at 10% by 2027 and 50% by 2047, and those aggregates have moved earlier in successive surveys.
Why does the calculator separate timing from severity?
Because they are different questions and fusing them produces nonsense. A countdown that speeds up when you raise a hostility slider is claiming that a hostile AI arrives sooner, which does not follow, and it implies that anything far enough in the future is automatically safe. Arrival date and outcome are modelled separately here for that reason.
Why does adding ten points to one slider matter more than another?
Because the four terms are multiplied. Ten points added to a term sitting at 15% nearly doubles the final answer; the same ten points on a term at 85% changes almost nothing. In a product the smallest link dominates, which is why the tool names it for you. It is usually the thing you are actually arguing about.
Are the preset numbers what those people actually said?
The headline figures are: the survey aggregates, and Carlsmith's roughly 5% by 2070. The individual conditionals are not, because nobody publishes those separately. They are one illustrative way of reaching the published total, and the tool says so under each preset.