# 2028–2032 - diffusion and the institutional lag

← [Part IV](README.md) · [Index](../README.md)

---

Capability gains continue, but **the story shifts to deployment.** The interesting variable stops being what models can do and becomes how fast organizations, regulators, professions, and underwriters can reorganize around what they already do.

The rate-limiting factor is **institutions and organizations.** That claim is the backbone of the middle period, and it is the main thing [Uncertainty 1](../06-uncertainties/recursive-self-improvement.md) would falsify - if research cycle times compress hard enough that the technology outruns the friction rather than waiting on it. Three governors still sit on that outcome (verification, physical supply chain, financial), only one of them technical; the fast-takeoff scenarios remain over-weighted relative to their full constraint set.

## What happens

### 1. Measured productivity finally moves

The J-curve resolves - the intangible-capital investment of the preceding years starts showing up in the statistics, roughly on the schedule that electrification and ERP both followed. The lag was never about model capability; it was about workflow redesign, and workflow redesign runs at the speed of management turnover. → [Game 4](../02-games/4-labor.md), [Part V row 2](../05-probabilities/)

"Speed of management turnover" is a mechanism, not a metaphor. The electrification precedent is specific: factories owned electric motors for two decades while keeping the line-shaft floor plan, because the people who had built their careers on the old layout were still running the plants; the productivity jump came when a new managerial cohort rebuilt factories around distributed drive. The AI equivalent is the difference between *installing* a copilot and *reorganizing* a function around the assumption that first drafts are free and verification is the scarce input. The first is procurement and shows up in month one; the second changes spans of control, junior pipelines, and what gets reviewed by whom, and it is resisted by exactly the people whose expertise the old workflow encodes. Firms founded after the capability existed skip the redesign cost entirely, which is why sector-level statistics move when cohort replacement does - entrants displacing incumbents, not incumbents transforming. Failure mode: if incumbent adoption this cycle is genuinely faster than the precedent - because the interface is language, not capital equipment - the turnover mechanism overstates the lag, and row 2's 2030 number is too pessimistic. The discriminator is whether productivity dispersion *between* firms in the same sector widens (redesign story) or stays flat while the mean rises (frictionless-tool story).

Two caveats that will make the statistics hard to read:

- **Quality-adjusted deflation is largely invisible.** A large share of consumer surplus lands as free tiers, bundled features, and time savings that never touch a transaction. Measured real growth understates true welfare gains. → [Prices](../09-macro/prices.md)
- **The METR-shaped perception gap may still be operating.** Self-reported productivity remains unreliable in a known direction. Weight instrumented and revealed-preference measures over surveys. → [Indicators B](../07-indicators/diffusion/README.md)

### 2. Robotics reaches commercial viability in structured environments

Warehouses, ports, agriculture - not general-purpose, not unstructured homes or hospitals. The structured/unstructured gap is the most reliable prediction in the robotics section: structure means repetition, repetition means samples, and samples are the entire constraint. → [Robotics](../03-domains/physical/robotics/)

Teleoperation-to-autonomy ratios, not demo videos, are the indicator. Any deployment reporting impressive capability without disclosing the ratio should be read as reporting teleoperation. → [Part VII headline indicators](../07-indicators/)

This is also when [logistics](../03-domains/physical/logistics.md) and [agriculture](../03-domains/physical/agriculture.md) stop being speculative domains and start being where the cost-curve claims meet payroll.

### 3. The first major AI-attributed incident occurs

And the regulatory architecture of the 2030s is written in its aftermath - within about **18 months** of it. Coordination is event-driven, not reason-driven; that is the historical pattern for aviation, nuclear power, finance, and pharma alike. → [Game 2](../02-games/2-nations.md)

The 18-month figure is a base rate, not a guess: Sarbanes-Oxley passed nine months after Enron's collapse; Dodd-Frank about twenty-two months after Lehman; the FAA's grounding-and-reform cycle after the 737 MAX crashes ran on a similar clock. The window exists because political salience decays - after roughly two years, an incident stops being a mandate and becomes a talking point, so whatever text is mature enough to move inside the window is what passes, drafting quality be damned. Two corollaries follow. First, the *quality* of the 2030s regime was mostly determined before the incident, by what was sitting on the shelf. Second, the regime will be shaped to the incident's specific mechanism - a finance-triggered event produces model-risk rules that say little about biosecurity, and vice versa - which means the modal outcome is a regime well-fitted to the previous war. The failure mode of the base rate itself: those precedents are single-jurisdiction responses; an incident whose harm crosses borders may produce the *domestic* architecture on schedule while the international layer stays empty, which is exactly the split [Part V row 4](../05-probabilities/) prices.

The mechanism is more likely [finance](../03-domains/cognitive/finance.md) or a correlated cyber event than [warfare](../03-domains/contested/warfare.md); the latter is less probable and less recoverable. [Biosecurity](../03-domains/contested/biosecurity.md) sits between them in probability and above both in tail severity.

The governance work that mattered *before* this window was having the good architecture drafted and on the shelf. By this window, the shelf is either used or it isn't.

### 4. Insurance, not legislation, sets the near-term deployment frontier

[Liability for AI systems is priced by underwriters](../03-domains/cognitive/insurance.md) before it is settled by courts. The refusals matter more than the coverage: fully autonomous action in high-severity domains, and anything with **correlated failure across many insureds at once**, get excluded or capped. Where AI is uninsurable, it does not get deployed by any organization with a board, regardless of capability.

This is a harder constraint than regulation, arrives earlier, and is set by people with money at stake. The correlation problem - few foundation models, synchronized update cycles, shared failure modes - is the real one, and it is why reinsurance and eventually a state backstop become standing proposals after the first correlated event. → [Uncertainty 6](../06-uncertainties/correlated-risk.md)

### 5. The apprenticeship gap becomes an acknowledged crisis

Roughly the point at which the 2026–28 hiring decisions have compounded into a visible seniority shortfall. Professional bodies, large firms, and governments start treating junior pipelines as strategic rather than as cost - or they don't, and the commons failure runs longer. → [Uncertainty 3](../06-uncertainties/apprenticeship-gap.md)

[Demography](../09-macro/demography.md) reframes the stakes without resolving them. Aggregate displacement is easier to absorb in a shrinking labor force; the *compositional* claim is untouched and possibly worse - a smaller entering cohort competing for a shrunken number of training positions produces the same 2040 expert shortage from both directions. **The apprenticeship gap is a larger share of total harm than the pre-demography document implied.**

### 6. The two-economy price split becomes the political fact

**Deflation in anything cognition-intensive; inflation in energy, land, healthcare, and skilled physical trades.**

That divergence is the defining political-economy fact of the period. It is the direct macroeconomic expression of [Game 3](../02-games/3-firms.md) - cheap cognition, expensive complements - and it will be experienced by most people not as "AI is deflationary" but as "everything I actually need costs more while everything I can download is free."

The split is **regressive on its face**: low-income households spend disproportionately on the inflating basket. Gains are real, widely distributed, and easy to under-notice; losses land in the most salient monthly expenses. **This is more consequential for 2030s politics than the employment story.** → [Prices](../09-macro/prices.md), [Assets](../09-macro/assets.md)

Central banks face one instrument and an index averaging two opposite movements. Monetary policy against supply-constrained inflation is close to counterproductive. Expect energy-price interventions, housing-supply politics, and pressure on healthcare costs - described as reactions to AI when they are reactions to prices.

### 7. Regional paths diverge on different constraints

| Region | What this period tests |
|---|---|
| [US–China](../03-domains/contested/geopolitics/bipolar.md) | Whether US permitting or Chinese chip access binds first; whether the compute-governance trap is acknowledged |
| [India](../03-domains/contested/geopolitics/india.md) | Whether 10M workers/year can be absorbed outside the exposed export sector |
| [Europe](../03-domains/contested/geopolitics/europe.md) | Whether regulatory cost falls on European adopters while advantage accrues to foreign producers |
| [Gulf](../03-domains/contested/geopolitics/gulf.md) | Whether capacity purchases convert into an economy, or remain a depreciating reserve swap |
| [Global South](../03-domains/contested/geopolitics/global-south.md) | Whether free technology reaches anyone without the complements - connectivity, credit, electricity, tenure |

**Welfare and growth come apart**, especially in the Global South: people living better while their country's development strategy fails is the modal outcome on this framework, and most analysis still conflates the two.

## What does *not* resolve here

- **The Taiwan assumption still carries this window.** A blockade or conflict in this window removes the majority of leading-edge capacity for years; there is no near-term substitute. → [Uncertainty 4](../06-uncertainties/taiwan.md)
- **Learned verification does not yet retire the master asymmetry** - or if it does, this is when the corpus ages poorly all at once. → [Steelman](../08-method/steelman.md) §1, [Uncertainty 5](../06-uncertainties/learned-verification.md)
- **Humanoids at >1M units/yr** is still a later-period question. Structured autonomy crosses commercial thresholds; general-purpose does not. → [Part V row 5](../05-probabilities/)

## What to watch

| Signal | Meaning | Indicator |
|---|---|---|
| TFP and labor-productivity prints, quality-adjusted where possible | Whether the J-curve is resolving | [B-family](../07-indicators/diffusion/README.md) |
| Teleoperation-to-autonomy ratios in deployed robots | Real robotics progress | Headline indicator |
| First salient AI-attributed incident + legislative response speed | Which governance architecture we get | [C-family](../07-indicators/governance.md) |
| AI-specific liability as a published insurance line | Deployment frontier being priced | [Insurance](../03-domains/cognitive/insurance.md) |
| Entry-level : senior ratios through 2029 | Whether institutions are responding to the gap | [Uncertainty 3](../06-uncertainties/apprenticeship-gap.md) |
| Relative price of cognitive vs. physical services | The two-economy split becoming measurable | [Prices](../09-macro/prices.md) |
| Grid interconnection queue length | Whether the energy ceiling is loosening | [A4](../07-indicators/substrate.md) |

## Failure modes

- **If diffusion looks like consumer software rather than electrification** - bottom-up, self-serve, no retooling cost - the institutional-lag argument is too long by 3–5 years and every date in this period should shift earlier. That is [steelman](../08-method/steelman.md) §3, and METR replication is the empirical hinge.
- **If the incident does not arrive by 2031**, Game 2's event-driven coordination claim weakens and the "no agreement before accident" prediction needs a logged revision.
- **If reinsurance capacity for AI liability materializes cleanly**, the correlation problem was overstated and deployment runs ahead of the insurance-constraint story.

### This period is where the two-economy split becomes political

[B7](../07-indicators/diffusion/economy.md) can print for years as a statistical curiosity. In this window it becomes a **household experience** - software and cognitive services soft, shelter/energy/care hard - and therefore a political fuel for [C5](../07-indicators/governance.md)-style backlash channels that are not "AI policy" at all. Housing and ratepayer politics are the AI distribution politics of the early 2030s whether or not anyone brands them that way. → [prices](../09-macro/prices.md), [assets](../09-macro/assets.md)

**Evidence family: statistical + legal.** TFP, insurance lines, price split, first incident - not train-run announcements or biped demos. Wrong family is a category error.

---

**Related:** [Prices](../09-macro/prices.md) · [Insurance](../03-domains/cognitive/insurance.md) · [Geopolitics](../03-domains/contested/geopolitics/) · [Game 2](../02-games/2-nations.md) · [Indicators](../07-indicators/)

**Next:** [2032–2040](2032-2040.md) · **Previous:** [2026–2028](2026-2028.md)