Extrapolating public data with a Poisson model dates the first AI catastrophe at March 2028

Stevekaplanai · reddit · 2026-09-19

A Reddit user built a simple model from public data estimating the median arrival of the first AI catastrophe at March 2028 (80% interval: mid-2027 to late 2028).

Inputs:

Method: a nonhomogeneous Poisson process — incident rate grows with deployment; the catastrophic fraction grows with capability.

Listed assumptions: 7.7 destructive incidents/year baseline, doubling yearly; 1% go catastrophic today, doubling with capability; "catastrophic" = $1B+ damage or critical-infrastructure disruption. Slower-growth scenarios push the date to late 2028, not never.

Original post →

More from AGI Musings

AGI Musings channel →