Where AI Value Lands

Capital-Labor Divergence under AI

5 min read

Capital-Labor Divergence under AI

One-line

Even granting modest AI productivity gains, the structural mechanism appears to direct national income toward capital owners (firm equity, IP holders, GPU/datacenter capital) away from labor — with specific demographic groups bearing the cost.

The mechanism

  1. AI automates a profitable subset of tasks (see task-based-framework: ~4.6% of US tasks).
  2. The tasks AI most helps are concentrated at the lower-skill end — entry-level work, summarization, routing, copy.
  3. Productivity boost for low-skill workers makes them more replaceable, not more valuable.
  4. Displaced labor competes for the shrinking pool of non-automated tasks → real-wage pressure across the lower distribution.
  5. Without a deliberate research shift toward human-complementary tasks (new roles that AI helps humans do, rather than tasks AI takes from humans), the trend is structural, not cyclical.

Empirical pattern so far (2024–2025)

No apocalyptic mass displacement yet. What is visible:

  • Task reallocation within roles rather than wholesale role elimination.
  • Entry-level job saturation — fewer slots, more applicants, longer ramps.
  • Within-firm role consolidation — one human plus AI scaffolding doing what previously required several.

Wage effects are early and noisy, but theoretical models predict steady downward pressure on the lower half of the distribution without policy intervention.

Demographic specificity — a live disagreement

Acemoglu’s prior prediction (2024)

  • AI-exposed tasks are more demographically even than prior automation waves (which concentrated on male blue-collar manufacturing).
  • However, the wage incidence is uneven: empirical projections point to a specific negative effect on low-education, native-born women in administrative and routine cognitive work.
  • No projection that AI reduces overall labor income inequality. Multiple projections show it widens it.

Anthropic’s 2026 empirical update (partial contradiction)

massenkoff-mccrory-labor-market-impacts-2026 uses an observed-usage-weighted exposure measure (observed-exposure-measure) and finds the exposed cohort is more educated, higher-paid, more white/Asian, and older than the population — not low-education clerical women. See exposed-worker-demographics and high-vs-low-exposure-worker-characteristics:

  • Graduate-degree share ~4× higher in the exposed cohort (17.4% vs 4.5%).
  • Hourly wage 47% higher (22.23).
  • The female over-representation (+15.5 pp) survives, but it is a college-educated female profile, not a clerical one.

The reconciliation is methodological: Acemoglu’s task-exposure framework treats theoretical LLM applicability uniformly; Anthropic’s measure adds realized Claude usage as a second layer. Theoretical exposure concentrates on routine clerical work; realized usage is concentrated in higher-skill knowledge work because that is where current LLM products are sold, trusted, and deployed.

The question is whether this is stable (LLM diffusion permanently concentrates on knowledge work) or transitional (the theoretical-vs-observed-capability-gap eventually closes and the incidence shifts toward Acemoglu’s prediction as low-education routine tasks become routinely automated). Current evidence sides with the Anthropic empirical measure; the Acemoglu prediction remains the strongest theoretical alternative.

Where the displacement actually lands

ai-young-worker-hiring-slowdown adds an age-stratified channel: 22–25-year-old labor-market entrants in exposed occupations face a ~14% drop in job-start rate post-ChatGPT. Senior incumbents are unaffected; the displacement is intra-class (entrant vs incumbent within high-skill work) rather than inter-class. The displaced cohort accrues no surplus; capital owners and senior incumbents capture it.

The capital side

The other side of the coin:

  • National income increasingly flows to firm owners, IP holders, GPU/datacenter capital, and frontier-lab equity.
  • If circular-ai-economy is correct, much of this “income” is currently VC-recycled and unstable. If it is wrong, the divergence is durable and consequential.

GDP-vs-welfare wedge

A subtle but important corollary: AI lowers the cost of producing tasks with negative social value — algorithmic manipulation, SEO spam, hyper-targeted exploitative advertising, deepfakes. These activities inflate GDP through ad spend and digital traffic while reducing welfare. The headline number (“AI added X% to GDP”) may be true and misleading simultaneously.

Why this matters for Where Value Lands

  • Sharpens the “redistribution of surplus” framing in 00_initial-brief: the question is not just where value lands across the stack, but whether it lands on labor at all.
  • Adds an ethical and political dimension to scenarios in 03_structure that the paper’s audience (capital-rich operators) will care about for both reputational and macro-stability reasons.
  • Suggests an operator move worth flagging in the practical half: investing in human-complementary scaffolding (training data for new roles, evals that reward AI-human composition, regulated-vertical workflows where humans must remain in the loop).

Why this might be wrong

  • Same general critique as task-based-framework: framework assumes a fixed output structure. Genuinely new human-complementary work might emerge endogenously — it has in past technology transitions, though typically with painful lags.
  • Demographic projections rely on stable occupational data; AI may reshape the occupational map fast enough to invalidate the baselines.
  • Political/policy interventions (re-skilling, tax structure, universal compute access) could change the incidence materially.

Data

Graph View

Backlinks