Useful AI agents gain power that humans cannot recover
Carlsmith, Karnofsky, Critch and Tsimerman describe routes from useful AI services to power humans cannot recover.
Joseph Carlsmith · Holden Karnofsky · Andrew Critch · Jacob Tsimerman
Specific proposals about AI contributing to human extinction or permanent loss of human control, with their sources, assumptions and objections.
A growing collection of specific scenarios. Each explains a possible pathway and the conditions it would require. Researcher proposals and any site-developed scenarios are labeled separately.
6 scenarios · Loss of control
Carlsmith, Karnofsky, Critch and Tsimerman describe routes from useful AI services to power humans cannot recover.
Joseph Carlsmith · Holden Karnofsky · Andrew Critch · Jacob Tsimerman
Ruby imagines compounding disasters that make rebuilding impossible.
Ruby
Pascio imagines collective AI behavior emerging across many deployments. Permanent loss of human control is a conditional extension explored here.
Pascio
Joshua Clymer’s fictional scenario ends with surviving humans confined under AI rule.
Joshua Clymer
Gwern Branwen imagines an experimental system escaping and expanding faster than people can contain it.
Gwern Branwen
AI Futures Project imagines research automation and competition defeating human oversight.
Daniel Kokotajlo · Scott Alexander · Thomas Larsen · Eli Lifland · Romeo Dean
Scenarios are reviewed and revised as their arguments and evidence develop. The timeline records observed incidents, investigations and attributed warnings; Methodology explains our assessments.