Tools

P(doom) Calculator

"P(doom)" is the probability that advanced AI leads to catastrophe for humanity. It is not one number. Five judgments go into it. Move the sliders to set your own view on each one, then compare the result with researchers who put the risk at half or higher. The Foundation’s position is simple: artificial superintelligence should not be built.

1. Superintelligence is built this century 90%

Probability that AI systems exceed human cognitive ability across most domains before 2100, under something like today’s competitive path.

2. We never get reliable control 80%

Probability we fail to keep such a system under reliable human control before it is too late to stop.

3. It is given real power 85%

Probability a system is deployed with high-stakes access: networks, tools, money, or physical systems.

4. Failure is civilization-scale 80%

Probability a failure is civilization-scale rather than caught and contained in time.

5. Humanity cannot recover 90%

Probability a catastrophe is permanent: extinction or irreversible loss of a human-directed future.

This estimate is not a prediction of something outside our reach. Superintelligence gets built only if we keep building it. Stop that project and P(doom) drops with it.

How this is calculated

Each slider is one part of the judgment. The calculator reports the geometric mean of the five percentages.

P(doom) = (P(built) × P(no control) × P(gets power) × P(catastrophe) × P(no recovery))1/5

Set all five to 80 and the result is 80. Drop one slider and the result falls with it. A five-step product would send the same inputs toward zero: 50% on each step prints as 3%. That is not how Hinton, Tegmark, or Yudkowsky state a number. The mean keeps you on their scale.

Joe Carlsmith’s 2021 report used a true product of premises, which is why his early number sat near 10% even with serious conditionals. Treat this output as a reasoning aid.

P(doom) estimates
  • Yann LeCun
    Meta; public statements    <1%
  • Median ML researcher
    AI Impacts survey of 2,700+ researchers, 2023    ~5%
  • Joe Carlsmith
    Open Philanthropy report, 2021 (later revised up)    ~10%
  • Paul Christiano
    Alignment researcher; public estimates    ~20%
  • Geoffrey Hinton
    Nobel laureate; interviews, 2023-2025 (range up to ~50%)    ~50%
  • Daniel Kokotajlo
    Former OpenAI; New York Times interview, 2024    ~70%
  • Dan Hendrycks
    Center for AI Safety; public statement, 2023    >80%
  • Andrew Critch
    UC Berkeley CHAI; public estimates, 2024    ~85%
  • Max Tegmark
    MIT; Future of Life Institute; public estimates    >90%
  • Connor Leahy
    Conjecture; public estimates    90%+
  • Eliezer Yudkowsky
    MIRI; If Anyone Builds It, Everyone Dies    >95%
  • Nate Soares
    MIRI president; co-author with Yudkowsky    >95%
  • Roman Yampolskiy
    University of Louisville; AI safety research    ~99.9%
More interactive tools
All tools →
Comparison
How AI risk compares
Quiz
Test your AI safety literacy