Catastrophe risk modeler: Tohoku tsunami shows historical data misleads on AI risk

davidmanheim · x · 2026-09-16

David Manheim, who spent years building catastrophe risk models, uses tsunami risk as an analogy for AI catastrophic risk:

Unlike rogue AI, tsunamis have real mortality data enabling falsifiable long-horizon predictions, e.g. global cumulative tsunami deaths staying below 10,000 over 2026–2035, provided Pacific and Indian Ocean early-warning systems keep operating through a major subduction-zone quake.
The Tohoku tsunami shows why purely historical-data estimates — rather than fundamental/mechanistic models — are often misleading (Fukushima was designed to historical frequencies and was overwhelmed).
He cites his 2018 paper criticizing overconfident pandemic-risk estimates, and notes the same methodological failure keeps recurring; the COVID response failures make one line from that paper haunt him.

Core claim: historical base rates are insufficient for assessing low-frequency, high-impact risks, including AI.

More from AGI Musings