Risk Probability
What Is P(Doom)? The Number People Use for AI Extinction Risk
P(doom) sounds scientific because it is written like a probability. Usually it is not the output of a repeatable experiment. It is a person’s judgement about the chance that advanced AI leads to catastrophe, often extinction or irreversible loss of control.
WHAT DOES “DOOM” MEAN TO A HOUSEHOLD?
A probability on a podcast feels abstract. A dead grid, an empty pharmacy and a child asking when the water returns do not. The useful horror question is not whether someone’s p(doom) is 5% or 50%. It is what the bad outcome actually contains.
Turn the number into consequences: failed control, broken logistics, war, subjugation or extinction. Then preparation stops being an argument and becomes a list of dependencies.
Scenario: this is a deliberately extreme “what if?” exercise, not a claim that these events are happening or certain to happen.
The immediate answer
P(doom) is informal shorthand for “probability of doom”. In AI discussions it usually means someone's personal estimate that advanced AI causes a catastrophic outcome. There is no official p(doom), and different speakers may define “doom” differently.
Why the numbers are all over the place
The categories above describe viewpoints, not calibrated probabilities. Nobody has a historical dataset of superhuman AI takeovers from which to calculate a frequency.
What is actually useful about p(doom)?
It forces people to expose how seriously they take the tail risk. If someone says the probability is effectively zero, ask which link in the catastrophe chain they believe cannot happen. If someone gives an extremely high number, ask which capabilities and deployment conditions they expect to arrive, and why countermeasures fail.
The useful conversation is in the assumptions underneath the number.
What p(doom) does badly
- It compresses many different catastrophe pathways into one number.
- It hides disagreements about definitions.
- It creates false precision.
- It can turn uncertainty into a personality contest between optimists and pessimists.
- It tells a household almost nothing about what practical resilience actions make sense.
Our rule
This site will discuss published forecasts when they are relevant, but it will not present one person's p(doom) as “the probability AI kills us”. The honest headline is that expert views vary enormously and the evidence is still developing.
How to read a p(doom) claim
First ask what “doom” means. Extinction? Permanent loss of political control? A global catastrophe with survivors? Then ask the time horizon. A one per cent estimate over five years and a one per cent estimate over a century describe very different beliefs. Finally ask what assumptions drive the number: rapid capability growth, poor alignment, weak governance, deliberate misuse or some combination.
If those details are missing, the percentage is more heat than light.
Do not prepare to a percentage
A household does not need to decide whether p(doom) is two per cent or twenty per cent before storing water, protecting accounts or making an offline communication plan. Prepare for consequences you can actually affect. Leave the grand probability argument to forecasters, and make your own resilience decisions using cost, usefulness and how many different emergencies the same preparation covers.
Why the debate gets emotional
People are not only arguing about mathematics. They are arguing about trust in technology companies, confidence in governments, faith in future safety research and different attitudes toward low-probability disasters. That is why the same evidence can produce radically different personal estimates.
Bayesian language does not make a forecast objective
People sometimes describe p(doom) as a Bayesian belief: a personal probability updated as new evidence arrives. That can be intellectually useful, but the result still depends on starting assumptions and judgement calls. There is no meter measuring the true probability of an unprecedented future event.
The number therefore works best as a summary of a person’s worldview, not as a statistic equivalent to an insurance-loss rate or weather probability.
Better questions than “what is your number?”
Ask which capabilities worry the forecaster, what deployment conditions they assume, which safeguards they think will fail and what evidence would make them revise their view. Those questions expose the model underneath the percentage.
For practical preparedness, the most useful question is simpler: which disruptions are plausible enough, cheap enough to prepare for and serious enough that a little redundancy pays off?
Turn the percentage into assumptions
A p(doom) number becomes more useful when it is unpacked into the beliefs underneath it. Does the speaker expect rapid capability growth? Do they think powerful systems will receive broad autonomy? Do they believe alignment or containment methods will fail? Are they worried mainly about deliberate misuse, accidents or autonomous loss of control? Two people can quote the same percentage while imagining completely different pathways.
The time horizon matters just as much. A personal estimate over the next decade is not comparable with an estimate spread across the rest of the century unless that difference is made explicit. Definitions matter too: some people use “doom” for extinction, while others include permanent political disempowerment or a catastrophic but survivable global disaster. Without those details, the neat number hides more than it reveals.
For a household, none of this produces a precise shopping list. The sensible route is to prepare for consequences that are useful across many scenarios: temporary loss of power, communications, payment access, trusted information or supply. Those preparations can make sense even for somebody who assigns an extremely low probability to the most severe AI outcomes.
Numbers need context
A percentage without a definition, time horizon and underlying assumptions tells you far less than its precision suggests.
Continue from here
Follow the scenario into the systems and household preparations most likely to matter next.
Evidence desk
These sources are used to separate demonstrated capabilities from the catastrophe scenario being explored.
Evidence review: August 2026 · P(doom) is treated as subjective forecasting, not measured risk.