Generative AI both replaces and enhances tasks across U.S. occupations, but the gains are uneven: higher‑credential jobs capture most augmentation dividends while middle‑skill workers face outsized displacement risk—an estimated 9.1 million worker equivalents are exposed to substantial displacement pressure.

Main Finding

The paper develops an AI Dual-Track model that maps task‑level exposure of 846 U.S. occupations to two divergent AI outcomes—Substitution (displacement) and Facilitation (augmentation)—using an ensemble LLM classification of O*NET competencies. It documents an “exposure paradox”: AI-driven displacement is widespread across the occupational distribution, but augmentation gains and their economic returns are concentrated among occupations requiring higher formal education. This pattern—termed digital decoupling—means access to AI tools alone is not the primary divider; rather, institutional credentials and facilitation literacy determine who converts AI exposure into productivity and economic gains. The study estimates that over 9.1 million worker-equivalents in middle-skill occupations face substantial displacement pressure, while higher-education groups capture the largest net labor-capacity gains.

Key Points

• AI Dual-Track framework: distinguishes Substitution (High-AI-Potential) and Facilitation (Moderate-AI-Potential) pathways at the task/competency level, producing a unified view of displacement vs. augmentation.
• Ensemble classification: 63 O*NET competencies were retained for modeling—28 classified as High-AI-Potential (substitution) and 35 as Moderate-AI-Potential (facilitation); 57 largely physical/psychomotor competencies were excluded as Low-AI-Potential.
• Educational stratification: Occupations are grouped by O*NET Job Zones (used as proxies for education levels). Higher Job Zones (bachelor’s and above) are more likely to realize facilitation/augmentation benefits; lower and many middle-skill occupations concentrate substitution risk.
• Exposure paradox & digital decoupling: Broad exposure to AI does not imply equal opportunity to benefit—the central labor-market divide is between those who can leverage AI to augment their work (often credentialed) and those whose tasks are more directly replaced.
• Policy framing: Introduces “educational immunity” (institutionalized cultural capital that shifts exposure toward facilitation) and “facilitation literacy” (capabilities needed to capture augmentation dividends) as targets for public policy to promote inclusive outcomes.
• Analytic scope caveat: The model maps structural exposure and latent capacity changes (Net Labor Capacity, NLC) rather than predicting adoption timelines or realized job-loss counts.

Data & Methods

• Data sources:
  • O*NET KSA descriptors (120 competencies; knowledge, skills, abilities) as primary task taxonomy.
  • BLS employment counts and annual wages by SOC codes.
  • Final analytic sample: 846 SOC occupations after merging O*NET and BLS.
• Education proxy:
  • O*NET Job Zones used as occupational-level proxies for formal education thresholds (1–5 scale mapped to no HS through master’s+).
• Competency classification:
  • Ensemble AI Classification Protocol using multiple LLMs (Gemini 1.5 Pro, ChatGPT-4o, Claude 3.5 Sonnet, Microsoft Copilot, Perplexity Pro).
  • Consensus reconciliation loop to reduce single-model idiosyncrasies; accounted for an “Accountability Gap” concept when assigning competencies to tracks.
  • Result: 28 competencies → Substitution track; 35 → Facilitation track; 57 Low-AI-Potential excluded.
  • Examples (from validated set): Displacement-linked competencies include Administration & Management, Programming, Operations Monitoring; Facilitation-linked include Communications & Media, Education & Training, Social Perceptiveness, Originality.
• Outcome metric:
  • Net Labor Capacity (NLC) constructed by weighting occupational competency importance with substitution/facilitation designations across Job Zones (paper frames NLC as a structural exposure measure; details and appendices provide full tables and prompting logs).
• Limitations acknowledged:
  • Ex‑ante structural/simulation framing — not a time-series adoption forecast.
  • Use of Job Zones as an occupational proxy for education (no individual-level microdata).
  • Classification depends on ensemble LLM judgments; although consensus was sought, this is not a ground-truth human coding.
  • Does not model firm adoption frictions, regulatory responses, labor supply adjustments, or macro stabilization.

Implications for AI Economics

• Distributional dynamics and skill premium:
  • AI may increase the skill premium by concentrating augmentation rents in higher-credential occupations, intensifying wage and productivity divergence unless policy intervenes.
  • Middle-skill occupations face measurable displacement pressure (paper cites ~9.1 million worker-equivalents), implying potential labor-market polarization absent countervailing mechanisms.
• Measurement and modeling:
  • Task-level, multi-track exposure metrics (substitution vs. facilitation) should be integrated into empirical models of technical change (e.g., task-based SBTC models) rather than relying on binary exposure or occupation-level aggregates.
  • Net Labor Capacity (NLC) or similar structural indices can help decompose productivity gains into machine-driven vs. human–AI complementarity components.
• Policy and institutional responses:
  • Education and credentialing matter for who captures augmentation gains—policies should aim to broaden access to facilitation literacy (curricula, public training, credential portability) rather than only private upskilling.
  • Active labor-market policies (targeted retraining for middle-skill workers, portable benefits, stronger social insurance) and occupational monitoring can mitigate displacement risks.
  • Design regulation and procurement policies to reduce algorithmic stratification (e.g., bias audits for career tools, incentivize workplace AI that complements rather than substitutes).
• Research priorities:
  • Link occupation-level structural exposure to microdata (worker outcomes, wages, transitions) to estimate realized effects and elasticities.
  • Model dynamic firm adoption, complementarities, and equilibrium labor supply responses to evaluate timeline and magnitude of displacement vs. augmentation.
  • Validate LLM-based competency classifications with human subject-matter coding and longitudinal outcomes; refine facilitation vs. substitution weights empirically.
• Broader economic interpretation:
  • The results suggest AI-driven productivity growth is not automatically inclusive; without public stewardship of facilitation capabilities and credential access, AI could reinforce existing inequalities and generate enduring digital decoupling between aggregate productivity and broad-based labor income gains.

If you want, I can: - Extract the validated 63-competency table into a CSV or spreadsheet. - Draft an outline for incorporating NLC into an empirical wage equation or task-based production function.