A method for reading the AI decade.
Most commentary on AI is unidirectional: one cause, one effect, one villain, one fix. Each analysis is roughly correct and beside the point, because the conditions do not act one at a time.
This framework is a discipline for seeing the conditions together: published indicators, a small number of analytical moves, a register of the mechanisms that couple one pressure to another, and a public record of what we predicted and how it turned out. The workings are exposed so they can be checked and revised.
The thesis
AI is an accelerant acting on societies whose conversion machinery was already failing.
By conversion machinery we mean what turns effort into standing: a degree into a job, savings into a home, a household into a family, work into meaning. In much of the developed world these conversions had been weakening for decades before generative AI existed. Each weakening is survivable alone. They do not occur alone. They compound, and institutions built to manage one variable at a time cannot see the compounding.
What AI changes is the speed, and it acts on the weakest link first: the entry level, the young. Employment among 22-to-25-year-olds in the most AI-exposed occupations fell 16 per cent relative to trend after the systems became generally available, while older workers in the same occupations were unaffected. The bottom rung was removed through hiring cessation rather than dismissals (Stanford Digital Economy Lab, ADP payroll data, November 2025).
Two properties govern everything. The conditions multiply rather than add: the graduate who cannot get the job also cannot leave the childhood bedroom, and neither statistic describes that life. And the geometry is not symmetric. Prevention is cheap. Cure may not be available at any price.
None of this makes an outcome inevitable, which is the point of building the instrument. A society that can see its compounding pressures can act on them. We do not have to be passengers.
Why now
Three things changed together. The capability became real enough to move labour-market data rather than conference agendas. Access to frontier capability turned out to be a permission, not a law of nature, and permissions can be revoked. And the political environment into which both land is already stressed: trust thin, housing unresolved, demography turning. The same technology arriving into a resilient society would be a productivity story.
The seven moves
The durable core of the framework. Indicators will be added and retired; mechanisms will be proposed and struck out; the moves should outlast all of it.
Move 1
State, not story.
Before narrating an event, locate it. A headline is a reading on an instrument, and the first question is which instrument. Much news moves nothing.
Move 2
Level or rate.
Does the development change where a society is, or how fast it is moving? The two have different governing variables and different remedies. AI is mostly a rate variable.
Move 3
Multiply, don't add.
Ask which existing condition a development compounds with, and name the mechanism. Youth underemployment times housing cost is one household that never forms. No named mechanism, no claimed connection.
Move 4
Check the asymmetry.
Is the change reversible, and at what cost? Trust, household formation, professional ladders and legitimacy run down fast and rebuild slowly. Margin banked early compounds in a society's favour.
Move 5
Watch the valves and the binds.
Valves are exits that vent pressure, above all whether the young can leave. Binds are the stakes that tie people to the existing order: property, family, a ladder worth climbing. Both are measurable.
Move 6
Trace the conversion chain.
Effort, credential, job, home, family, meaning. A link that stops converting produces the aggrieved. Effort that stops signifying produces withdrawal, which is worse and harder to see.
Move 7
Fuel or spark.
Structural pressure is measurable. Ignition is not. Measure fuel load, decline to predict sparks, and require a control case for every claimed pattern.
Meta
The weights are unknown, and we say so.
Anyone offering a fixed model of how these conditions combine is describing a religion. The indicators are real. The weights are a published prior and nothing more.
Fear and agency
A framework about compounding pressure will sometimes be frightening, and it should be. The reader is owed the fear the facts support: specific and sourced. What the reader is never owed is fatalism. Doom says the trajectory is fixed. Hype says the technology guarantees a good outcome. Both are forecasting errors, and both tell the reader to go back to sleep. Every reading ends at the same question: what moves the distance from here, and who can move it?
What this is built from
The components have a long lineage and we cite it. Structural-demographic theory (Goldstone, Turchin) supplies elite overproduction and the fiscal-legitimacy spiral. Hirschman supplies exit and voice. Merton supplies the difference between the aggrieved and the withdrawn. Gurr supplies relative deprivation. Systems dynamics supplies stocks, flows, feedback and hysteresis. The synthesis is ours. The parts are not.
What is new:
- AI treated as a rate variable, with technical fluency. The lineage models had no term for a technology that compresses decades of diffusion into years.
- Operationalisation at news cadence, with public scoring. We run the instrument weekly, publish what we expect, date it, and score it in public.
- A mechanism register with controls. Every claimed interaction has a named pathway, an evidence status, and a control case.
- The middle-power vantage. The instrument is built from the position of democratic middle powers, whose dependencies are part of their condition.
Claim classes
Every forward-looking statement carries one of three labels:
- Structural: how the machinery works. Slow-moving, defended by mechanism and lineage, revised rarely.
- Conditional: if these conditions co-move, then this pressure builds, by this mechanism. Testable against the panel.
- Forecast: dated point predictions, scored in public. A forecast stated as a certainty is a register violation.
What would count against this framework
Stated in advance, because a framework that cannot specify its own failure conditions is a mood:
- Mechanism failure. A registered mechanism's conditions co-move and the predicted effect repeatedly fails to appear. The mechanism is retired in public.
- Rate failure. If AI-exposed labour markets diverge favourably for new entrants over a sustained period, the accelerant term is wrong as stated.
- Prediction failure. If our forecasts resolve at or below chance over a meaningful run, the framework is decoration and we say so.
- Index failure. If rank ranges widen until the number carries no information, the index is retired and the panel is published without composites.
Revision protocol
The framework is versioned like software. Weights, thresholds, indicators and mechanisms change through a recorded revision: what changed, why, what evidence forced it, effective date. Nothing is corrected silently.
Last updated