AI safety needs aviation's playbook: fatalities per quintillion tokens may drop 10x a year

robleclerc · x · 2026-09-15

Rob LeClerc argues AI won't be risk-free, and aviation history offers a sober template: zero risk was never the standard (aviation killed thousands yearly for decades), absolute and relative harm can move oppositely (AI deaths per quintillion token may fall 10x per year), accident-driven institutional feedback loops make systems safer, and society's harm tolerance ratchets up over time — early AI-era harms deemed normal in 2026-2030 may look unacceptable by 2040.

Original post →

More from AGI Musings

AGI Musings channel →