The Next Fifteen Years

A forecast built from first principles
Section future / 03-domains / contested / state-capacity.md

State Capacity - the precondition for every governance claim#


Contents

Every governance argument in this document assumes a state capable of acting. Game 2 predicts an architecture written in the 18 months after a salient accident. Energy predicts electricity politics arriving before capability regulation. Insurance predicts a state backstop after the first correlated event.

None of that happens well unless the state can hire, understand, and enforce. That capability is currently the weakest link in the entire chain, and it is almost never analyzed at the same resolution as model evals.

Four capacities, not one#

CapacityQuestionFailure mode
MeasurementCan the state independently verify claims about systems?Rules without inspection → theatre
TalentCan it hire or borrow people who understand the stack?Capture by default; industry drafts the regime
AdoptionCan it use AI on its own workloads?Private sector pulls away; public backlogs persist
EnforcementCan it impose costs on non-compliant actors across borders?Law on paper; leakage in practice

Most "AI governance" debate is about what the rules should say. This page is about whether any rule can be real.

1. Measurement - the first problem#

A regulator cannot enforce what it cannot measure. Right now, no government can independently verify:

Every one of those is currently established by self-report from the regulated entity. That is the structural situation in pharmaceutical approval before the FDA had inspection authority, and in financial reporting before independent audit - and both took decades and a scandal each to fix.

The governance investment that pays best is not rulemaking but measurement capability - the technical ability to check claims independently. Without it, any architecture written in the post-incident window is unenforceable, and unenforceable rules produce compliance theatre plus a false sense of resolution, which is worse than no rules at all.

Compute governance as verification (C3, bipolar trap) only works if someone can audit chips, clusters, or attestations. That is a measurement institution problem before it is a treaty problem.

2. Talent - stated honestly#

The state needs people who understand this technology deeply, and it is competing for them against employers paying an order of magnitude more, at a moment when the scarce input has opinions about where it works.

This isn't new - the same gap exists in financial regulation, aviation, and pharma - but the compensation ratio is more extreme and the technical half-life is shorter. Three partial mitigations, none sufficient alone:

Without talent, C2 shelf-readiness is empty: the post-incident window fills with whatever draft industry had ready.

3. Adoption - the irony#

Government is unusually well-suited to the current technology, and unusually badly positioned to adopt it.

Suited, because a large share of state activity is exactly the cheap-ground-truth cognitive work that compresses: benefits adjudication, permit review, tax examination, procurement analysis, translation, case backlog triage, records management. Backlogs are the state's characteristic failure mode, and backlogs are precisely what abundant cognition dissolves.

Badly positioned, because of procurement cycles measured in years, legacy systems measured in decades, civil-service rules that make redeployment slow, and an asymmetric error regime: a wrongly-denied benefit is a headline, a slowly-processed one is a statistic. That asymmetry rationally produces extreme caution, and extreme caution is expensive when the alternative is a queue.

the largest realized public-sector gains through 2032 are in backlog clearance - permits, courts, benefits, immigration, veterans' claims - rather than in anything resembling policy analysis. Boring, unglamorous, and worth more than most of what gets announced.

The permitting loop#

Permitting throughput is a state-capacity question, and permitting is the binding constraint on energy, which is the binding constraint on AI itself. A state that used AI to clear its own permitting backlog would be relieving the constraint on the technology by deploying the technology. That is the one administrative intervention that pays for itself, and it requires no new statutory authority in many jurisdictions - only procurement and willingness.

Same logic applies to grid studies (energy sector), court dockets (law), and licensing boards that gate medicine.

4. Enforcement - borders and balance sheets#

Even a competent measurer faces:

Enforcement capacity is why insurance and liability often bind first: they do not need a new agency, only courts and underwriters. State capacity still matters for the backstop after correlated failure (Uncertainty 6).

Variance across states#

TypeMeasurementAdoptionImplication
High-capacity administrative states (parts of N. Europe, Singapore, etc.)Medium–highMediumCan run real regimes if they choose; politics may still refuse
High-tech, fragmented states (US federal)Uneven by agencyLow–mediumElectricity politics and state-level rules outrun federal capability policy
High-build, party-state (China)Different transparency problemHigh where prioritizedCapacity without liberal verification; bipolar asymmetry
Low-capacity / debt-stressedLowLowRules imported, unenforced; Global South lives with private governance

"Regulation arrives" is not a single global event. It is a distribution over state types. The corpus's US-centric Game 2 timeline is a claim about one high-salience jurisdiction, not a world average.

The uncomfortable part#

If the state cannot measure, cannot hire, and cannot adopt, then the Game 2 prediction needs a caveat it does not currently carry.

The architecture written in the post-incident window will be written by whoever has the technical capacity to draft it - which, absent state capability, means the regulated industry, standards bodies it funds, or a small number of civil-society organizations with fewer resources than either.

That is not necessarily a bad outcome; industry drafting produced workable regimes in aviation and finance. But it is a different outcome from the one implied by "regulation arrives," and it should be predicted explicitly rather than discovered later. Capture is not a risk to the process; under low state capacity, it is the default shape of the process.

The measurement institution needs a pipeline, not a mandate#

One refinement on the first section, because "fund a measurement institution" is the sort of recommendation that gets adopted in name. Measurement capability is not primarily people or authority - it is standing access to the thing being measured, arriving continuously rather than on request. An inspectorate that can demand documents after an incident is doing forensics; one that receives training-run telemetry, deployment logs, and evaluation results as a matter of routine is doing measurement. The difference is the same one that separates a financial regulator with reporting feeds from one with subpoena power, and the first is what makes the second usable.

That reframes the hard part. The obstacle is not statutory authority, which is comparatively easy to legislate, but the engineering and confidentiality machinery to receive sensitive operational data from competing firms without leaking it or being captured by whoever operates the pipe. The institutions that solved this - clearing-house reporting, aviation safety reporting, and nuclear materials accounting - each took years to build the plumbing and had a scandal in their history that funded it. Failure mode: the plumbing is also the capture surface. An agency dependent on a voluntary feed from the entities it regulates has a strong institutional interest in not disrupting the relationship, which is the well-documented failure of every self-reported regime this one would be modeled on.

What would change this#

Indicators#

SignalReading
Independent eval / red-team capacity funded inside governmentMeasurement layer real or not
Time-to-hire for technical roles in AI-relevant agenciesTalent gap
Permit / docket clearance rates where AI tools deployedAdoption working
Post-incident rule text authorship (agency vs industry consortium)Capture vs capacity
Interconnection queue movement after "AI for government" programsPermitting loop closed or rhetorical

Soft-cap note: this page is near the expansion ceiling; further depth should split (measurement vs procurement vs talent) rather than pad. The indicators table is the operational core - if only one section is maintained, maintain that.


Related: Game 2 - Nations · Energy constraint · Energy sector · Law · Uncertainty 2 · Governance indicators · Europe as rule-setter vs builder

View markdown source

select · Enter open · Esc close