A skills-focused simulation suggests AI’s technical reach in the U.S. labor market is far larger and more geographically widespread than visible adoption indicates: current AI capabilities overlap with about 11.7% of U.S. wages (~$1.2tn), compared with roughly 2.2% (~$211bn) reflected in present adoption concentrated in tech hubs.

Main Finding

Project Iceberg introduces the Iceberg Index, a skills-centered metric that quantifies the wage-value share of occupational tasks AI systems can technically perform. Using Large Population Models to represent 151 million U.S. workers (923 occupations, 3,000+ counties) and a catalog of 13,000+ AI tools mapped to 32,000+ skills, the baseline Index shows that visible AI adoption in computing/tech (the “surface”) accounts for ~2.2% of wage value (~$211B), while technical capability beneath the surface spans administrative, financial, and professional services, totaling ~11.7% (~$1.2T) of wage value—about five times larger and geographically widespread. The Index measures exposure (capability overlap), not displacement or timing.

Key Points

• What the Iceberg Index measures
  • Skills-centered KPI: percent of wage value in an occupation tied to skills AI can technically perform.
  • Focus: technical exposure (where human and AI capabilities overlap), not forecasts of job loss or adoption timing.
  • Construction: weights each skill by importance, automatability (tools + LLM tool-use), and prevalence to produce 0–100% exposure per occupation.
• Scale and scope
  • Human population: 151M workers, 923 occupations, 32,000+ skills, 3,000+ counties.
  • AI population: 13,000+ tools (copilots, workflow automation, etc.).
  • Simulation engine: Large Population Models implemented in AgentTorch, run on Oak Ridge Frontier supercomputer.
• Key quantitative results
  • Surface Index (observable tech-sector adoption): 2.2% of wage value (~$211B).
  • Iceberg Index (baseline technical capability across economy): 11.7% of wage value (~$1.2T).
  • Traditional metrics (GDP, income, unemployment) explain <5% of cross‑regional variation in skills-based exposure.
• Validation
  • Skill embeddings vs. career transitions: 85% recall for predicting commonly observed occupation-to-occupation moves, supporting that skill-based representations capture real labor-market structure.
  • Adoption alignment: 69% geographic agreement between the paper’s Surface Index and Anthropic Economic Index (AEI) usage tiers.
• Uses
  • Baseline assessment of maximum technical exposure (no adoption assumptions).
  • Input to scenario simulations that model adoption speed, transferability, and policy interventions.
  • Policymaker applications: identify exposure hotspots, prioritize training/infrastructure, reassess multiplier assumptions, test interventions before large investments.

Data & Methods

• Data sources and representations
  • Occupational skill taxonomy (32k+ skills) mapped from standard sources (e.g., O*NET) to individual workers and AI tools.
  • Worker population modeled as agents with attributes: skills, tasks, location, occupation.
  • Catalog of 13k+ AI tools mapped to same skill taxonomy to assess tool availability for tasks.
• Model & computation
  • Large Population Models (LPMs) simulate billions of interactions among workers, skills, and tools; implemented in AgentTorch.
  • Simulations run at national scale on Oak Ridge’s Frontier supercomputer.
• Index computation
  • For each occupation, exposure = sum over skills of (skill importance × automatability × prevalence), normalized to 0–100%.
  • Automatability determined by evidence that tools exist and language models can use them to perform the skill in at least one context.
  • Baseline Index reports the maximum technical exposure (i.e., capability demonstrated in any context) rather than expected adoption.
• Validation approach
  • Structural validation: skill-embedding similarity compared to observed career transition networks (85% recall).
  • Adoption validation: compare Surface Index rankings to AEI real-world usage tiers (69% agreement; high agreement at extremes).
• Limitations noted in methods
  • Current baseline focuses on digital/cognitive tool exposure; physical/robotic automation excluded for now.
  • Index does not model firm adoption decisions, regulatory effects, worker acceptance, or labor-market dynamics that determine realized outcomes.

Implications for AI Economics

• Measurement and forecasting
  • Existing labor statistics (employment, wages, GDP) miss substantial AI-mediated exposure; skills-centered metrics are required to anticipate where capability exists before observable adoption.
  • The Iceberg Index provides a forward-looking input for scenario analysis and cost–benefit calculations of training or infrastructure investments.
• Policy and workforce strategy
  • Many exposure hotspots lie outside coastal tech hubs—states and regions should not equate low current AI employment with low exposure.
  • Policymakers can use the Index to prioritize reskilling where high-wage-value skills are automatable, to design targeted transitions (e.g., administrative automation freeing clinical time), and to reassess job-multiplier assumptions under increased automation.
  • The platform enables pre-deployment testing of interventions (training programs, incentives, regulations) to evaluate effectiveness before large expenditures.
• Research and market implications
  • Firms and training providers can identify transferable skill pathways and anticipate changes in entry-level vs. experienced hiring demand.
  • The large “hidden” exposure suggests substantial market for AI-augmentation tools in non-tech sectors (finance, admin, professional services).
• Cautions for interpretation
  • Exposure ≠ displacement: technical capability does not imply immediate job losses or reduced employment; outcomes depend on adoption, complementarities, policy, and worker adaptation.
  • Baseline depends on tool catalog and mapping assumptions; results should be updated as tools, LLM capabilities, and tool-use practices evolve.
• Directions for future work
  • Incorporate physical/robotic automation and capability-benchmark data (e.g., task-level LLM benchmarks) to refine automatability.
  • Use the baseline as input to dynamic adoption simulations to estimate realized labor-market impacts under alternative policy or market scenarios.
  • Extend geographic and firm-level analyses to evaluate sectoral multiplier shifts under different adoption pathways.

Summary takeaway: The Iceberg Index reframes AI’s labor-market impact as a skills-level exposure problem—revealing a much larger and geographically dispersed potential for AI to perform valuable work than is visible through current tech-sector adoption metrics—and provides a scalable simulation framework for policymakers and firms to test strategies before committing large investments.