Grace et al. 2023 survey of 2,778 AI authors: median 5 percent for extremely bad outcomes, October 2023

The document exists. The ledger fetched it at its publisher and quotes it.

Record ID

ODDS-2026-0021

Kind

Survey

Jurisdiction

Global

Last verified

16 September 2026

Added

16 September 2026

• The paper, Thousands of AI Authors on the Future of AI (arXiv 2401.02843), reports 2,778 responses from 18,459 working addresses, a 15 percent response rate, fielded 10 to 24 October 2023.
• The median prediction for extremely bad outcomes, such as human extinction, was 5 percent.
• Over a third of participants, 38 percent, put at least a 10 percent chance on extremely bad outcomes.
• The abstract: between 38 and 51 percent of respondents gave at least a 10 percent chance to advanced AI leading to outcomes as bad as human extinction, depending on the question.
• Of net optimists, 48 percent gave at least a 5 percent chance of extremely bad outcomes; 59 percent of net pessimists gave 5 percent or more to extremely good outcomes.

Role at the time, Where it was said, Timeframe named, Stated basis, Later revision. Stated means the document you can open below says it; silent means the ledger read the document and it does not.

Role at the timeStated

Survey by Katja Grace, Harlan Stewart, Julia Fabienne Sandkuhler, Stephen Thomas, Ben Weinstein-Raun, Jan Brauner and Richard KorzekwaGrace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, 5 January 2024.

Where it was saidStated

Fielded 10 to 24 October 2023; paper submitted to arXiv 5 January 2024, revised 8 October 2025Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, 5 January 2024.

Timeframe namedStated

Long run outcomes, not otherwise boundedGrace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, 5 January 2024.

Stated basisStated

Self reported credences from a 15 percent response sample of authors at six venuesGrace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, 5 January 2024.

Later revisionOpen

No later revision found as of 2026-09-16

1. 5 percent

   Median probability of extremely bad outcomes such as human extinction

   Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, as of 24 October 2023.

2. 38 percent of respondents

   Share of participants giving at least a 10 percent chance of extremely bad outcomes

   Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, as of 24 October 2023.

3. 51 percent of respondents

   Upper share, across question framings, giving at least 10 percent to outcomes as bad as extinction

   Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, as of 24 October 2023.

4. 2,778 responses

   Responses received

   Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843), primary source, as of 24 October 2023.

The same median twice, a year apart, with four times the respondents, tells you the middle answer is stable among those who reply; the spread between 38 and 51 percent across framings tells you how much the question wording moves the result. A reader should treat the median as a fact about respondents and the framing effect as a warning about every single number on this ledger.

1. Grace et al., Thousands of AI Authors on the Future of AI (arXiv 2401.02843)Primary · 5 January 2024
2. Grace et al., Thousands of AI Authors on the Future of AI, HTML versionPrimary · 8 October 2025

• VerifiedSurveyODDS-2026-0020

  AI Impacts 2022 expert survey: median 5 percent for an extremely bad outcome such as human extinction, August 2022

  Global · verified 16 September 2026

  The survey page: approximately 4,271 researchers who published at NeurIPS or ICML in 2021 were invited; 738 responses were received, some partial, a 17 percent response rate.

• Open questionOpen questionODDS-2026-0022

  Open question: are two named people's percentages comparable when neither states a timeframe or a definition of catastrophe?

  Global · verified 16 September 2026

  Grace et al. 2023 report that between 38 and 51 percent of respondents gave at least 10 percent to outcomes as bad as extinction depending on the question asked, a framing effect inside one survey.

• VerifiedReport findingODDS-2026-0003

  International AI Safety Report 2026: likelihood, nature and timing of loss of control unusually ambiguous, February 2026

  Global · verified 16 September 2026

  Page 76 of the report reads that a key challenge for policymakers is preparing for a risk whose likelihood, nature, and timing remains unusually ambiguous; the ledger read the sentence in the PDF at arXiv 2602.21012 and at the report's own site.

• Open questionOpen questionODDS-2026-0023

  Open question: when a frontier lab researcher states a personal probability, what does it tell you about the lab?

  Global · verified 16 September 2026

  Hubinger wrote that he personally thinks the chance is greater than 10 percent within the next decade and that Anthropic does not yet have a plan to solve alignment for superintelligence (ODDS-2026-0001).

• VerifiedSigned statementODDS-2026-0018

  Future of Life Institute: pause for at least six months the training of systems more powerful than GPT-4, March 2023

  Global · verified 16 September 2026

  The letter at futureoflife.org, dated 22 March 2023: we call on all AI labs to immediately pause for at least 6 months the training of AI systems more powerful than GPT-4.

• VerifiedSigned statementODDS-2026-0017

  Center for AI Safety Statement on AI Risk: extinction risk a global priority alongside pandemics and nuclear war, May 2023

  Global · verified 16 September 2026

  The statement text at safe.ai: mitigating the risk of extinction from AI should be a global priority alongside other societal scale risks such as pandemics and nuclear war.