The Next Fifteen Years

A forecast built from first principles
Section future / 07-indicators / governance.md

C - Governance Indicators#


Contents

Baselines as of mid-2026. Cadence: event-driven.

Game 2 argues that coordination is event-driven rather than reason-driven: the architecture governing the 2030s gets written in the 18 months following a salient accident. This family therefore has an unusual structure - most of it sits still, and then all of it moves at once.

The purpose of watching it during the quiet period is to know which draft is on the shelf when the window opens.

C1 - The incident#

BaselineNo AI-attributed event large enough to force a legislative cycle
Assumed window2027–2031
Most likely mechanismA market dislocation - Finance has the fastest loops, the least human latency, and the most correlated positioning
Least recoverableWarfare - an escalation event has no rollback
RevisesPart V row 3; everything in family C downstream of it

Watch for near-misses, which are more informative than the eventual hit and arrive earlier: cyber intrusions with autonomous lateral movement, biological synthesis screening failures, flash dislocations traced to correlated model-driven positioning. A near-miss that gets publicly attributed is a small version of the forcing event and produces a small version of the regulatory response.

One distortion to guard against: attribution is itself contested terrain. Whether an event counts as "AI-attributed" is decided politically after the fact, by actors with positions to defend - which means C1 can fire late (a genuine AI incident laundered into operator error) or early (an ordinary failure branded as AI to serve a regulatory agenda). The indicator tracks the public attribution, because that is what opens the legislative window; the analyst should separately track whether the attribution was earned, because that determines whether the resulting architecture addresses a real mechanism.

C2 - Shelf-readiness of the response#

BaselineFragmented; no consensus architecture drafted and defensible
Trigger - preparedA coherent, technically literate regime published and endorsed across factions before the incident
Trigger - unpreparedThe window opens with only advocacy positions available
RevisesThe quality, not the timing, of 2030s governance

The most actionable item in the document. The window will be short and whatever is ready will win - this is the observed pattern in aviation, nuclear power, financial regulation, and pharmaceutical approval alike. Drafting quality now determines outcomes later far more than advocacy volume does.

C3 - Compute governance viability#

BaselineExport controls on accelerators; no verification regime for training
Trigger - viableOn-chip attestation or location verification deployed at scale
Trigger - deadFrontier capability routinely reached below any plausible reporting threshold
RevisesGame 2, Part V row 4

The lever weakens exactly as the stakes rise: algorithmic efficiency and inference-heavy methods both erode the correlation between "large training cluster" and "frontier capability," which is the correlation the entire compute-governance approach rests on. Track efficiency gains as an indicator of governance feasibility, not just of capability.

C4 - Diffusion rate of frontier capability#

BaselineOpen-weight tier ~3–6 months behind frontier (Epoch AI, Jan–May 2026; down from the ~9–15 months carried at authoring)
Trigger - controls workingGap widening past ~24 months
Trigger - controls failingGap closing below ~6 months
RevisesGame 1, Game 2

Leads decay at maybe 30–50%/yr absent extraordinary security. This measures whether that decay rate is being changed by policy or merely described by it. A lead you cannot hold is a commercial asset, not a strategic one - and the width of this gap is the difference between the two.

C5 - Electricity-price politics#

BaselineWholesale prices up sharply at datacenter-adjacent nodes; localized political reaction
TriggerLarge-load tariffs, siting moratoria, or ratepayer-protection statutes in two or more major markets
RevisesEnergy, 2028–2032

The argument here is that this arrives before any federal capability regulation does, because it is the only channel through which AI touches a median voter's budget monthly. If federal safety regulation lands first, that ordering claim was wrong and the political model in Game 2 needs revisiting.

C6 - Liability allocation#

BaselineUnresolved; contracts allocate by negotiation, courts have not settled it
TriggerA precedent-setting judgment or statute assigning liability for autonomous system harm
RevisesLaw, Medicine, Game 1, and every "human retains the part where someone can be sued" claim in Part III

Underrated because it is slow and unglamorous. But the accountability layer is the human moat in most of Part III, and its width is set entirely by where liability lands. A ruling that shifts liability onto model providers collapses that moat in several domains simultaneously; one that keeps it with the deploying professional entrenches it for a decade.

C7 - AI liability insurance capacity#

BaselineNo material AI-specific commercial line with published rates; cyber forms used as imperfect substitute; correlation poorly modeled
Trigger - private capacityDistinct AI liability line with published rate tables and growing limits
Trigger - correlation bindsAggregate limits / event exclusions tighten; reinsurer withdrawal; or first correlated multi-insured loss event
Trigger - state backstopTerrorism-reinsurance-style proposal enacted or seriously drafted post-event
RevisesInsurance, Uncertainty 6, 2028–2032 deployment frontier

The exclusions define the deployment frontier more tightly than capability demos. C6 is precedent; C7 is price and availability. Together they say whether the moat is legal doctrine or balance-sheet arithmetic.

C3b - Compute-governance trap (export controls vs verification)#

BaselineAccelerator export controls used as competitive instrument; no bilateral verification regime
Trigger - trap deepensTighter controls + accelerating domestic substitution / efficiency that shrink the controlled set's relevance
Trigger - trap acknowledgedOfficial linkage of control regimes to a future verification architecture (even aspirational)
RevisesBipolar, Part V row 4, C3

Round 6 named the mechanism: the lever and the future arms-control verification tool are the same object. C3 tracks technical viability; this row tracks whether policy is spending the option.

C8 - Provenance and synthetic-media infrastructure#

BaselineC2PA and similar standards exist; adoption uneven; easy to strip; platforms inconsistent
Trigger - infrastructure realMajor platforms + device makers ship default capture-time attestation and consumer-visible trust UX at scale
Trigger - theatreLabeling mandates without attestation (easy spoof, high compliance cost, low trust effect)
Trigger - enclosure pathVerified content retreats almost entirely into walled gardens; open web treated as unverified by default
RevisesGame 5, Media, Compressed, B6, B11

Game 5 and media argue authenticity becomes a product and that open unverified content goes lemons. C8 is whether society builds the public good (point-of-capture provenance) or only private gardens. Differs from B6 (assessment reversion in education/hiring): C8 is content and identity infrastructure for the feed and the courtroom.

Watch legislation that mandates labels without requiring hard attestation - that is the privacy-class failure mode applied to media (rules without measurement). → State capacity

Near soft cap: prefer editing existing C-rows over adding new ones; a future split would separate incident/architecture (C1–C4) from liability/provenance (C5–C8) if growth continues.


Related: Game 2 - Nations · Part V - Probabilities · Warfare · Finance · Insurance · Media · Game 5

View markdown source

select · Enter open · Esc close