Daniel Kokotajlo

Former OpenAI governance researcher; lead author of the 'AI 2027' scenario; executive director of the AI Futures Project

AI Futures Project

Safety researchers

Kokotajlo walked away from millions in OpenAI stock rather than sign a gag clause, then wrote 'AI 2027', a widely read scenario of AI automating AI research within a few years. He puts the chance of things going 'horribly wrong' around 70% and now expects superhuman coding around 2028.

Their number

70%

"roughly a 70% probability of ending in catastrophe, including but not limited to human extinction" — he clarifies this covers loss of control, AI takeover, authoritarian concentration and war, not only extinction (Diary of a CEO, August 2026; the 70% figure dates to his 2023 LessWrong comments).
Source

On timing

AI 2027 scenario (April 2025) depicted superhuman coders in 2027. In the Q1 2026 update his median for an 'automated coder' moved to "mid 2028" (from late 2029); superintelligence median around 2029.

Source

What they call for

  • 'right to warn' whistleblower protections
  • Transparency about frontier capabilities
  • US China agreement to slow frontier development ('Plan A')
  • Compute supply chain audits and inspections
  • Ability to disable frontier compute if needed

In their own words

  • Co-signed 'A Right to Warn about Advanced AI': risks range "to the loss of control of autonomous AI systems potentially resulting in human extinction"; "AI companies have strong financial incentives to avoid effective oversight."

    4 June 2024

    Source

  • Left OpenAI refusing a non-disparagement agreement, forfeiting roughly $2 million in equity.

    2024-05

    Source

  • Q1 2026 update: automated-coder median "mid 2028" (Daniel), "mid 2030" (Eli Lifland).

    2 April 2026

    Source

Signed

  • Right to Warn letter 2024

Last verified 1 September 2026

Report a correction

Back to all voices