The Next Fifteen Years

A forecast built from first principles
Section future / 06-uncertainties / correlated-risk.md

Uncertainty 6 - Correlated-failure insurability as the deployment frontier#


Contents

Insurance argued that underwriters price AI risk before legislators do, and that exclusions - not coverage - define where AI actually deploys. The variance that belongs in this register is narrower and sharper: whether correlated failure is insurable at any price the real economy will pay.

If it is not, the deployment frontier is a reinsurance and sovereign-backstop problem, not a capability problem. That would make every Part III timeline too slow or too fast depending on how states respond - and the corpus had been treating insurance as a domain page rather than a global constraint.

The structural problem#

Insurance works when losses are independent. AI risk is correlated by construction:

This is cyber risk's shape, which the industry already handles badly (NotPetya, MOVEit, CrowdStrike). AI concentration is worse: fewer effective providers, deeper integration, faster propagation.

Why the variance is high#

Three branches, with very different consequences for the rest of the document:

BranchWhat happensEffect on timelines
Private capacity formsAI-specific liability becomes a material line with published rates; aggregate limits bite but deployablePart V insurance-linked claims vindicated; deployment tracks capability more closely
Reinsurance is the ceilingA handful of reinsurers set the economy-wide deployment capSlower than capability; geographic and sectoral rationing of autonomy
State backstop after first correlated eventTerrorism-reinsurance template legislated in the 18-month windowBinary: either unlocks deployment or loads it with political conditions

Round 0 had no view on which branch. The insurance page (round 2) predicts branch 1 at ~70% by 2030 for the existence of a line, while flagging that correlation could still force branch 2–3 for the high-severity tail. The open uncertainty is the tail, not the product category.

The second-order claim that makes this matter#

If AI liability is expensive to insure, the inelastic complement in most domains is not expertise - it is the balance sheet that can absorb residual risk.

That reframes human moats from capability claims into financial ones, and financial arrangements change on an underwriting cycle, not a training cycle. It is the most plausible mechanism by which Part III timelines turn out far too slow - or, after a backstop, far too cautious.

Leading indicators#

SignalReading
AI-specific liability as published commercial line with rate tablesBranch 1 forming
Aggregate limits and event exclusions tightening at renewalCorrelation being priced in
Reinsurer public commentary / withdrawal from AI-adjacent cyberBranch 2 binding
Model-provider indemnification breadth in enterprise contractsWhether risk stays with labs or spills to insureds
Legislative proposals for AI terrorism-style backstopBranch 3 preparing, usually post-event

Anatomy of the first correlated event#

The adversarial case is sharper when the event is specified rather than gestured at. The ingredients are all standing conditions: a synchronized update pushed to every deployment at once, agentic write-access to production systems rather than draft output, and a failure mode subtle enough to run for days before detection. The CrowdStrike outage of July 2024 is the dry run - one vendor, one update, millions of machines, and insured losses widely estimated at a small fraction of the economic loss, most of which fell into exclusions and retentions. The AI version differs in one respect that matters to underwriters: CrowdStrike failed loudly and instantly, which capped the loss window at hours. A model-driven failure that corrupts decisions rather than crashing machines has a detection lag measured in days or weeks, and loss scales with the lag. Terrorism insurance is the other anchor: private capacity withdrew overnight after September 2001 and returned only under the TRIA backstop - which is why the branch table above treats post-event withdrawal-then-backstop as a template with precedent, not speculation.

The pre-event pricing problem cuts both ways, and this is the part the corpus should hold with the least confidence. With no loss history, insurers either overprice (deployment slower than capability justifies, branch 2 arriving early and unnecessarily) or underprice (capacity looks abundant right up until the first event vaporizes it, branch 1 appearing confirmed and then failing in a week). Both errors generate years of indicator readings that look like a resolved uncertainty. A published rate table is only weak evidence for branch 1 until it has survived an event; the indicator that actually discriminates is renewal behavior after the first loss, and there is no way to observe it early.

The coupling to complement erosion#

Uncertainty 7 lists liability capacity among the slowest-eroding complements, and this page is why: a balance sheet that can absorb correlated residual risk is not manufacturable by intelligence at any capability level - it is manufacturable only by accumulation or by sovereign fiat. If this uncertainty resolves to branch 2, the balance-sheet row becomes the most durable complement on the list and the distributional predictions in assets concentrate further (reinsurers and states as the ultimate complement-owners). If it resolves to branch 3, a legislative act erases the scarcity overnight - the same political repricing risk U7 flags for licenses. The two pages disagree productively: U7 says political scarcity is the least durable kind; this page says the political act requires a triggering catastrophe, which puts a floor under the timeline.

What would falsify the constraint story#

Relationship to other uncertainties#

What "fires" looks like in the data#

Three signatures, in rising severity. (1) Exclusion language spreads from specialty lines into ordinary professional liability and cyber - the quiet form, already partly present. (2) Aggregate limits and co-insurance appear on AI-assisted work even where coverage nominally exists - capacity without usable limits is capacity in name only. (3) A correlated loss event that hits multiple insureds on the same model generation - the event that converts a theoretical argument into a reinsurance market freeze. C6 and the insurance page's product watch are the leading indicators; waiting for (3) to score this uncertainty is waiting for the resolution event itself.

Monoculture is the underwriting input. Herding on the same model family, scaffold, and update schedule is what turns independent professional risk into catastrophe risk. Watch model-share concentration in deployed enterprise stacks the way finance watches factor crowding - finance monoculture table applies here with different labels.


Related: Insurance · Game 2 · Game 1 on liability · Governance indicators · 2028–2032

View markdown source

select · Enter open · Esc close