Methodology
What is estimated. Each row is the chance, by 2100, of a specific failure — named by what goes wrong, not by a generic “doom” label. Most rows use the Existential Risk Persuasion Tournament outcome: more than 10% of humanity dying within about five years. The two AI rows instead count permanent disempowerment: Carlsmith’s outcome for loss of control, and lock-in (Forethought; Kulveit et al.) for power concentration. Because the outcomes differ, rows are comparable as “chance this specific bad future happens”, not as identical death tolls.
How each risk is built. Each risk is a chain of premises multiplied together. Every premise after the first is conditional on the one above it, so the product is the chance that the whole chain happens; the exact multiplication is shown under each category. Setting a premise to 0% switches off everything that depends on it. AI is split into the two mechanisms the scenario literature separates: rogue systems / loss of control (Carlsmith; Hendrycks et al. 2023) and power concentration (Forethought coups and Kulveit et al. gradual disempowerment, which share the precondition that AI replaces the human workforce). AI misuse for bioweapons and cyberattacks ends in a pandemic or grid failure, so it is a stated input to those premises rather than a separate row.
Where the defaults come from. Only the loss-of-control chain follows a published decomposition: Carlsmith's "Is Power-Seeking AI an Existential Risk?" (2022, to 2070). Its defaults keep his figures for alignment difficulty, power-seeking and escalation, but use 85% for systems being built by 2100 and 60% for coordination failing, against his 65% and 80%. The power-concentration chain is grounded in named scenarios — Forethought (2025), Kulveit et al. (2025), Christiano (2019) — but none publishes a probability, so its defaults are illustrative assumptions, flagged as such; treat that row as “what would have to be true”, not an expert estimate. The other chains were assembled for this site; their defaults were picked so each result lands inside the published range for that risk, for example nuclear catastrophe at 6% against 4–8% in the Existential Risk Persuasion Tournament. Figures under "What others say" marked "whole-chain risk" are estimates for the whole risk, not for that single premise, and are there for comparison. Note also the trade-off the single-chain model hid: making alignment easier lowers the loss-of-control row but can raise the power-concentration row, because obedient AI is a better tool for seizing power.
Combined chance. Calculated as 1 − (1 − first risk) × (1 − second risk) × … across the five rows, which treats them as independent. They are not: shared premises push several risks up together, and one event can count twice, such as a nuclear war that also collapses the grid. Both effects mean the combined figure probably overstates the chance of at least one catastrophe.
Ranking prep. Each action has a weight for each risk: 2 if it helps a lot, 1 if it helps, 0.5 or 0.25 if it helps a little. The weights are this site's judgement, drawing on Ready.gov, Finland's 72-hours programme, the Johns Hopkins Center for Health Security, Boyd and Wilson (2023), Benjamin Todd, Zvi Mowshowitz and John-Clark Levin. "Chance of a catastrophe where this helps" is 1 − (1 − risk × weight ÷ 2) multiplied across the risks the action covers: roughly, the chance that a catastrophic scenario the action helps with actually happens. Actions are ranked by that figure alone. Cost is shown but not used. Ordinary emergencies such as storms and local outages are not counted, though that is where most of these items get used. The figure measures how often an action is relevant, not how much it improves the odds of survival.