AI researcher Eli Lifland puts AI takeover risk at ~67% if labs don't slow down
eli_lifland · x · 2026-10-11
Prominent AI forecaster Eli Lifland laid out his current risk assessment:
- Timelines: AI companies could develop broadly superhuman AI (TED-AI — at least as good as top human experts at virtually all cognitive tasks) any year now: 10-15% by end of 2027, 50% by end of 2030.
- Takeover: If labs don't slow down, AI takeover probability is 67%.
- Extinction: Conditional on takeover, human extinction is 40%, and many non-extinction worlds still involve massive deaths.
In follow-ups, he cited his AI 2027-related reasoning, clarified that visible signs of misaligned takeover could come well after TED-AI (over 2 years later in AI 2027) or around the same time, and said he feels increasingly viscerally concerned about near-term catastrophic risk. He published subjective probabilities for more precise communication.
More from AGI Musings
- RSI without homeostatic constraints only accelerates drift, argues Indonesian techie — AryHHAry · 2026-10-11
- The Bitter Lesson 2.5 years on: why agreeing with it doesn't change your daily AI work — hyunw_kim · 2026-10-11
- JeffLadish: Pre-ASI AI can't break defense dominance, neither US nor China can win with it — JeffLadish · 2026-10-11
- 'Maths is over' is hubris: one mathematician says it has only just begun — timigod · 2026-10-11
- When AI can do the work, jobs shift from doing to judging what's worth doing — AryHHAry · 2026-10-11
- AI skeptic camp is self-contradictory: 'useless' yet career-threatening, argues creator — VraserX · 2026-10-11