Base Rates - Social Response (classes 3, 5, 6)#
Contents
How societies have historically absorbed a technology shock: when regulators move, what displacement actually does to the displaced, and the demographic tide the whole story plays out against.
3. Regulatory response cycles#
| Domain | Forcing event | Architecture written |
|---|---|---|
| Aviation safety | Successive fatal crashes | Within ~2 years of each |
| Nuclear power | Three Mile Island, Chernobyl | 18 months–3 years |
| Financial regulation | 1929, 2008 | 12–24 months |
| Pharmaceuticals | Elixir sulfanilamide, thalidomide | 12–24 months |
| Data privacy | Accumulated scandal, no single event | ~15 years, and weakly |
The pattern is event-driven, fast once triggered, and shaped by whatever was already drafted. This is the entire basis for Game 2's central claim and for the emphasis in C2 on shelf-readiness.
The privacy row is the important exception. Where harm is diffuse and gradual rather than concentrated and salient, no window opens at all and regulation arrives late, weak, and fragmented. If AI harms turn out to be diffuse - labor displacement, epistemic degradation, gradual dependency - the correct reference class is privacy, not aviation, and the Game 2 prediction is wrong in an important way: there is no forcing event and no architecture, just drift.
That is a live possibility and it is underweighted in this document.
What "shelf-readiness" actually buys#
The table's "architecture written within N months" column is not a claim that legislators invent rules from scratch after the crash. In every fast row, a draft was already sitting in a drawer - ICAO annexes, NRC rule packages, Dodd-Frank precursors, FDA statutes awaiting an incident to force votes. The window is for passage and implementation under political heat, not for first principles. That is why C2 tracks whether an AI architecture exists on the shelf before the incident: the reference class says the window is short enough that only pre-written work becomes law. A jurisdiction with no draft gets the privacy outcome even after a crash - frantic hearings, fragmented state action, and a decade of case law instead of a statute.
Mixture harms are the hard case. AI can produce both a salient incident (a market crash with models in the causal chain, a clinical autonomy failure, a cascade cyber event) and diffuse labor/epistemic harm in parallel. The class predicts the architecture written will address the salient channel and under-address the diffuse one - finance/cyber rules after a finance/cyber incident, not apprenticeship policy. Part V row 3 and row 4 are priced with that skew: incident probability is material; verified binding agreement is not.
5. Labor displacement episodes#
The China shock is the closest well-studied analogue, and the lesson taken from it here is specifically about concentration, not aggregate: aggregate employment effects were modest and locally the effects were severe, persistent, and politically decisive. Adjustment costs were far higher than trade models assumed because labor is not geographically or occupationally mobile on the timescale the shock arrives.
Two adjustments for the AI case, in opposite directions:
- Worse: exposure is occupational and cohort-based rather than geographic, so it hits people who cannot move away from it. And it lands on the entry rung specifically - see Game 4.
- Better: cognitive workers are more geographically mobile, better credentialed, and better capitalized than displaced manufacturing workers were. The absorption capacity is genuinely higher.
What the China-shock literature actually priced#
Autor, Dorn, and Hanson (and the follow-on political work) showed that the harm concentrated in commuting zones with high manufacturing exposure, persisted for a decade-plus, and rewired voting behavior even where aggregate unemployment later recovered. The mechanism was not "no other jobs exist" but local multipliers, skill mismatch, and non-migration: house prices, family ties, and information frictions kept people in places where the demand for their skills had collapsed. AI's version of that mechanism is occupational rather than geographic - a junior analyst in a coastal city cannot "move" out of the entry rung the way a worker could theoretically leave a mill town - which is why the concentration lesson transfers even when the mobility adjustment cuts the other way.
Political implication carried into priors: the force that writes rules will track visible cohort pain and media-salient incidents, not national unemployment rates. B1 (entry-level posting ratios, recent-grad unemployment) is therefore both a labor indicator and a political leading indicator. Aggregate unemployment staying low while B1 deteriorates is the China-shock pattern restated for knowledge work - and it is the pattern under which class 3's privacy exception becomes more likely, because diffuse cohort harm rarely produces the single forcing event aviation-style regulation requires.
6. Demographics - the counterweight (integrated round 7)#
Every advanced economy has a shrinking working-age population through the 2030s. Full treatment: Demography.
Result carried into the priors: aggregate displacement estimates are an upper bound; the apprenticeship-gap compositional claim is untouched and is a larger share of total harm than the pre-demography document implied. Adoption politics may be easier in labor-scarce economies than US-centric analysis assumes.
The demographic counterweight does not cancel class 5. It changes which harm shows up in the politics: fewer laid-off mid-career workers, more silent pipeline failure and wage pressure in the residual human roles. That is a quieter politics - closer to the privacy exception than to the aviation template - unless a care or housing crisis converts scarcity into a salient incident of a different kind.
Failure mode of this page: using "demographics help absorption" to dismiss B1. Compositional harm and aggregate employment are different objects; class 6 bounds the second and leaves the first untouched.
Mixture harms tip class 3. AI can produce both a salient incident and diffuse cohort harm. Class 3 predicts the architecture addresses the salient channel first - finance/cyber rules after a market event, not apprenticeship policy. Plan shelf drafts for both; expect only the salient one to pass.
Related: Game 2 · Game 4 · Demography · Uncertainty 3