Task-Level Gains Are Real but Economy-Wide Distribution Lags

AI boosts specific tasks significantly: GitHub Copilot speeds completion 55%, customer service productivity rises 14% (largest for novices per Brynjolfsson et al. 2023 study), coding tasks improve 20-45%, legal reviews 30-50% (Thomson Reuters). McKinsey projects $2.6-4.4T annual global value. Yet BLS shows only 1.7% labor productivity growth in 2025, above 2010-2019's 1.4% but below task predictions. Gains appear in profits (12% rise), S&P highs, executive pay, and AI infra like NVIDIA's market cap jump from $360B (2023) to $3T (2025). Median wages grew 0.8%; labor income share fell below 57%, extending a four-decade decline. Post-1973 decoupling means median workers earn ~$30K less yearly than if wages tracked productivity (EPI estimate). Historical IT boom (2.5% growth) and platforms created value but unevenly, favoring owners over workers.

AI's Cost Structure and Dynamics Favor Winner-Take-Most

Unlike past tech, AI has $100M+ training fixed costs but near-zero marginal inference (cents per query), enabling scale to billions of users and oligopolies. Network effects loop more data into better models, locking in leaders. Substitution (chatbots replacing agents, AI content/writing/code cutting headcount) sends gains to capital; augmentation could share them but firms opt for 10 developers doing 15's work via freezes, not raises. Knowledge work sees 20-40% boosts captured as reductions; manufacturing/creative sectors shift value to platforms/shareholders; services mix savings to consumers/profits over wages. Skill premium widens as AI aids high earners, displaces middle-skill.

Institutions and Choices Determine Sharing

Post-WWII (1947-1973) synced productivity/wages via 35% unions, 70%+ top taxes, full employment—unusual, not automatic. Eroded arrangements need replacement. Tax AI capital/windfall profits to fund retraining/transfers (tradeoff: slows innovation). Labor: sectoral bargaining, portable benefits, displacement disclosures. Antitrust: block AI mergers, mandate data portability. Public: open-source models, augmenting public AI. Tech leaders choose augmentation (build pro-AI constituencies) over replacement (invites backlash/regulation, as in EU). U.S. market-driven maximizes speed but risks narrow gains; Nordics/EU balance with protections. Deployment decisions shape decades of outcomes.