State Capacity - the precondition for every governance claim#
Contents
- §1 Four capacities, not one
- §2 1. Measurement - the first problem
- §3 2. Talent - stated honestly
- §4 3. Adoption - the irony
- §4.1 The permitting loop
- §5 4. Enforcement - borders and balance sheets
- §6 Variance across states
- §7 The uncomfortable part
- §8 The measurement institution needs a pipeline, not a mandate
- §9 What would change this
- §10 Indicators
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#
| Capacity | Question | Failure mode |
|---|---|---|
| Measurement | Can the state independently verify claims about systems? | Rules without inspection → theatre |
| Talent | Can it hire or borrow people who understand the stack? | Capture by default; industry drafts the regime |
| Adoption | Can it use AI on its own workloads? | Private sector pulls away; public backlogs persist |
| Enforcement | Can 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:
- The scale of a training run
- The capability of a deployed system
- Whether a stated safety evaluation was conducted as described
- Whether a model contributed to a specific harm
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:
- Rotational secondments rather than permanent hiring, accepting the conflict-of-interest cost as the price of competence
- Concentrated technical units with independent pay authority - how central banks and a few defense research organizations solved analogous problems
- Buying measurement rather than building it - funding independent third-party evaluation, which pushes the verification problem one layer out rather than solving it
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:
- Cross-border leakage - weights, talent, and inference APIs move; national rules face Game 2 diffusion
- Corporate structure - SPVs, cloud regions, and vendor chains blur who is the regulated entity (capital financing mix)
- Unequal counterparties - fining a startup is easy; compelling a hyperscaler or a sovereign lab is politics
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#
| Type | Measurement | Adoption | Implication |
|---|---|---|---|
| High-capacity administrative states (parts of N. Europe, Singapore, etc.) | Medium–high | Medium | Can run real regimes if they choose; politics may still refuse |
| High-tech, fragmented states (US federal) | Uneven by agency | Low–medium | Electricity politics and state-level rules outrun federal capability policy |
| High-build, party-state (China) | Different transparency problem | High where prioritized | Capacity without liberal verification; bipolar asymmetry |
| Low-capacity / debt-stressed | Low | Low | Rules 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#
- A well-funded, technically credible measurement institution - NIST-analogue for capability evaluation, with inspection authority and independent pay scales. Cheap relative to its leverage; currently the biggest gap.
- Procurement reform that lets the state buy software on software timelines. Repeatedly attempted, repeatedly stalled; precondition for public benefit from any of this.
- A visible backlog-clearance success. Nothing changes bureaucratic behavior like a peer agency demonstrably solving a problem everyone shares.
- Pay and status for technical civil service sufficient to staff the measurement layer - without which secondments remain the only path and capture remains structural.
Indicators#
| Signal | Reading |
|---|---|
| Independent eval / red-team capacity funded inside government | Measurement layer real or not |
| Time-to-hire for technical roles in AI-relevant agencies | Talent gap |
| Permit / docket clearance rates where AI tools deployed | Adoption working |
| Post-incident rule text authorship (agency vs industry consortium) | Capture vs capacity |
| Interconnection queue movement after "AI for government" programs | Permitting 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