AI safety researchers respond to 'how could AI kill everyone?' with five scenario write-ups
DavidSKrueger · x · 2026-09-12
After Jacob Coxon's resignation over extinction warnings, some claimed AI safety researchers have no realistic answer to 'how could AI kill everyone?' Researcher ohabryka compiled the five best scenarios he knows of:
- AI 2027 (Daniel Kokotajlo, Scott Alexander, et al.): predicts superhuman AI's impact over the next decade will exceed the Industrial Revolution, with detailed 'slowdown' and 'race' endings built on trend extrapolation and expert feedback; the authors stress 2027 was only a modal estimate, not a certain date.
- Paul Christiano's scenario (former head of safety at AISI, 2019)
- Gwern Branwen's scenario (independent AI researcher, 2022)
- Holden Karnofsky's high-level explanation (RSP Lead at Anthropic, 2022)
- Joshua Clymer's scenario
David Krueger (Cambridge AI safety researcher) shared the list for people asking the question, noting he doesn't endorse every scenario. The AI 2027 team is also offering prizes for the best alternative scenarios to spur debate.
Related event: AI Safety Feud Erupts Over Demand for a Concrete Extinction Scenario(7 posts)→
More from AGI Musings
- Ex-OpenAI researcher warns AI could self-copy across the internet, sparking backlash — rohanpaul_ai · 2026-09-12
- Bipartisan AI Workforce Commission Launches, Co-Chaired by Raimondo and Paul Ryan — sebkrier · 2026-09-12
- 25 Fields Medalists led by Terence Tao sign declaration against AI companies' math benchmarks — GaryMarcus · 2026-09-12
- Even if AI solves all Millennium Prize problems, skeptics will still dismiss it — haider1 · 2026-09-12
- tszzl pushes back on Fermi paper: 50-OOM lognormal abiogenesis prior under-justified — tszzl · 2026-09-12
- MIT Tech Review roundtable on Sept 15 asks: will AI really kill us all? — nordicinst · 2026-09-12