Game 4 - Labor: comparative advantage doesn't guarantee a wage#

← Part II · Index

Ricardo guarantees humans always have something to do. It guarantees nothing about the wage.

Horses had comparative advantage over tractors in some tasks and their population still fell 90%. The comparative-advantage argument is correct and reassuring about employment; it is silent on price. Those are different questions and they are routinely conflated.

The honest math#

• ~60% of US employment sits in occupations where ≥20% of tasks are meaningfully AI-exposed.

• But exposure ≠ substitution. RCT evidence is genuinely mixed: +14% for customer support agents, +25–40% for consultants on in-domain tasks, +26–55% in some coding studies - yet METR found experienced open-source developers were 19% slower with AI assistance while believing they were 20% faster.

That perception gap is the most important result in the literature and should make everyone humble. It means self-reported productivity data - which is most of the data - is unreliable in a known direction.

• Historical diffusion lag. Electrification took ~40 years to show up in productivity statistics because factories had to be physically redesigned around unit drive. ERP took ~10. AI ships over existing rails and self-serves, so it's faster - but the binding constraint is organizational redesign, not model capability, and org redesign runs at the speed of management turnover.

The aggregate estimate#

12–18% of US task-hours economically substitutable by 2032, of which maybe half is realized as headcount change - call it 10–14M job-equivalents over seven years. Against gross US job churn of ~5M/month, that's absorbable in aggregate and devastating in concentration.

Composition and concentration are the whole story#

The aggregate number is the least interesting part of it.

• Junior knowledge roles and middle-skill cognitive tasks - coordination, documentation, first-pass analysis, routine review - absorb the large majority of the pressure in this window.
• Physical, care, and high-stakes-liability roles absorb very little of it before ~2032. The constraint there is bodies and licensure, not intelligence.
• Geography compounds it. Exposure is spatially clustered in metros and regions built on a single exposed sector, exactly as manufacturing decline was.

Aggregate churn can absorb 10–14M job-equivalents. A single occupation, cohort, or metro absorbing its share of them within eighteen months cannot - not without the frictions (long unemployment spells, occupational downgrading, geographic immobility, political reaction) that made the China-shock literature what it is.

Forecasts that report only the aggregate are measuring the wrong object.

The sharpest harm is not mass unemployment. It's the apprenticeship gap.#

AI is best at exactly what juniors did: first-draft research memos, boilerplate code, document review, basic analysis, tier-1 support.

Firms rationally cut junior hiring. Each firm's decision is individually correct. Collectively, the profession has no seniors in 2040.

This is a textbook commons tragedy in human capital formation. It is already visible in law, consulting, and software hiring data. And almost no institution is set up to internalize it - the firm that trains juniors pays the full cost and captures a fraction of the benefit, which is the exact structure that guarantees undersupply.

The mechanism is worth stating in Becker's terms, because it explains why the problem got worse rather than merely arriving. Firms rationally fund specific human capital (useful mainly to them) and underfund general human capital (portable to competitors). Junior knowledge work used to bundle the two: the memo drafted for this client taught skills portable everywhere, and the firm paid for the bundle because it needed the memo. AI unbundles it - the firm can now get the memo without buying the training that rode along with it. The training was never the product; it was a positive externality of inefficient production, and efficiency destroyed it. That framing also says what a fix must look like: some institution has to buy the training as training, the way German-style apprenticeship systems, teaching hospitals, and military pipelines do - cost-sharing arrangements built precisely because on-the-job spillover cannot be trusted to survive cost pressure.

The cobweb problem: the price signal arrives too late#

Markets do contain a self-correction: if seniors become scarce in the 2030s, the senior wage premium rises, which raises the return to becoming one. But the production lag for a senior professional is on the order of a decade, so the market response has classic cobweb dynamics - today's hiring decisions respond to today's junior glut, the shortage price signal arrives around the time the missing cohort should already have been trained, and the correction overshoots into whatever the 2030s equilibrium is. Cohorts are not inventory; you cannot backfill 2028's missing associates in 2035. The prediction that follows, checkable by the early 2030s:

Senior-to-junior wage ratios in the exposed professions widen through the early 2030s, and firms respond less by reviving junior hiring than by compressing the seniority ladder - using AI plus a thinner layer of seniors, redefining "senior" downward, and poaching rather than training.

Failure mode: if AI closes the junior-to-senior gap faster than it eliminated junior work - juniors with AI performing at prior mid-levels, the Uncertainty 3 resolution - the cobweb never forms because the production lag itself collapsed. That is the strongest version of the counterargument, and it turns on whether judgment is learned by doing the automatable work or can be learned by supervising it.

Watch this indicator above all others: the ratio of entry-level to senior postings in knowledge professions. It is the leading edge of everything.

→ The case that this resolves rather than compounds is in Uncertainty 3.

What the 2024–26 data actually shows#

The apprenticeship-gap claim has now moved from prediction to partially-observed, which means it can be checked rather than argued. As of mid-2026:

The inversion is real and it is concentrated exactly where the model predicts: young, cognitive, exposed, and in the professions where the first two years of work were the most automatable.

But the causal attribution is genuinely contested#

Three confounds, and honesty requires stating that they are not resolved:

1. The rate cycle. Junior hiring is the most cyclically sensitive category in white-collar employment, and 2023–25 was a tightening cycle following a hiring bubble. Some of this is mean reversion from 2021 over-hiring, not substitution.
2. Remote work. A 2026 LSE study found remote-work prevalence a better predictor of declining entry-level hiring than AI exposure. The mechanism is plausible: distributed teams make the informal observation-and-correction that trains juniors far more expensive, so firms stop buying it.
3. Composition of job growth. Healthcare, government, and hospitality accounted for most net US job creation in 2024–25 - sectors with low AI exposure and low graduate intake. The graduate market can weaken without any AI channel at all if growth simply happens elsewhere.

This matters for the forecast, not just for the attribution. If the driver is cyclical, it reverses with the cycle. If it is remote work, it is a management-practice problem with known fixes. If it is substitution, it is structural and compounds. The three have completely different policy responses and completely different 2030s.

roughly 50% of the entry-level decline is AI substitution, 30% cyclical, 20% work-organization. That split is held with low confidence and is the single number in this document most likely to be revised.

The discriminating test runs through the next cycle: if junior hiring fails to recover when aggregate white-collar hiring does, the substitution share was large. That test resolves around 2027–28 and is worth more than any amount of further cross-sectional analysis. → Indicators

Comparative advantage does not pay a wage#

The classical reassurance - humans will specialize in what they are relatively better at - assumes that relative advantage still commands a market price above reservation utility. If AI absolute advantage covers both the junior task set and a growing share of senior judgment, the residual human tasks can be (a) high-skill and scarce (complements, good wages) or (b) low-skill and abundant (care, presence, bad wages) or (c) institutionally reserved (licenses). The distribution across (a)–(c) is set by Game 3's complement list and by demography, not by comparative-advantage slogans. This page's apprenticeship claim is about pathway (a) drying up at the entry point; demography and care shortages are about pathway (b) expanding. Both can be true at once.

50/30/20 is the weakest number on the page. Treat it as a prior to be destroyed by the 2027–28 recovery test, not as a stable parameter. If juniors recover with the cycle, rewrite toward cyclical/remote; if not, raise substitution and promote Uncertainty 3.

Related: Software and Law for where the gap is already visible · 2026–2028 · Uncertainty 3 · Indicators

View markdown source