Index 16% · Group

AI safety researchers

75%

Median of 6 counted estimates. One of 4 groups in the index, each with equal weight.

Who qualifies: People whose main work is AI safety or alignment: at a safety organisation, on a lab’s safety team, or in academia.

How the median is worked out

Counted estimates, lowest to highest. A range counts at its midpoint; estimates older than five years aren’t counted. With an even number, the median is halfway between the middle two.

  1. 10% Evan Hubinger
  2. 46% Paul Christiano
  3. 70% Daniel Kokotajlo
  4. 80% Dan Hendrycks
  5. 90% Max Tegmark
  6. 100% Roman Yampolskiy

Median: 75%

Stated estimates

Watchlist

People in this group who haven’t stated a number we can source. They join the list when they do.

  • Connor Leahy

    Listed elsewhere at 90% or more, from a video interview. We haven't yet confirmed his own words.

    Source

  • Eliezer Yudkowsky

    Often listed at above 90%, but those figures are other people's summaries of his views. We haven't found a number he has stated himself.

    Source

  • Geoffrey Irving

    Has described 'about a coin-flip chance' that everyone dies because of superintelligent AI. We list stated numbers, so this phrase isn't counted.

    Source

  • Holden Karnofsky

    Gave 10–90% as the range of views he finds reasonable, not as his own estimate.

    Source

  • Jan Leike

    Often listed at 10–90%, but in the cited interview that range was said by the host, not by Leike, who declined to give a number.

    Source

  • Toby Ord

    Gave about 1 in 10 for existential catastrophe from unaligned AI this century in his 2020 book The Precipice. We are confirming the exact wording.

    Source