AI-driven productivity is not self-sustaining: gains must reach household incomes or economies hit an 'absorption' ceiling that limits consumption and welfare improvements. The paper reframes the debate by linking technological efficiency to distributional transmission through wages, prices, investment and public policy and proposes measurable indicators for future empirical testing.

Main Finding

Webb develops an automated, general-purpose measure of occupational “exposure” to a given technology by measuring text overlap between patents and job-task descriptions, validates it on historical technologies (software and industrial robots), and then applies it to artificial intelligence (AI). He finds that: - Historical exposure scores for robots and software are correlated with subsequent within-industry declines in employment and wages. - AI patenting is concentrated on high-skilled, pattern-detection and judgment tasks (medical diagnostics, fraud detection, inspection, etc.), so AI exposure is highest among college-educated, older, and high-wage occupations. - If one assumes the historical linear mapping from exposure to wage changes continues to hold, AI exposure implies a modest reduction in 90:10 wage inequality (projected –4% using the software coefficient, –9% using the robot coefficient) but an increase in top-end inequality (99:90). Webb emphasizes these projections rely on strong assumptions and should be interpreted cautiously.

Key Points

• New, replicable exposure measure: uses patent text and job-task text to quantify where a technology is directed (task-level overlap).
• Method of overlap: extract verb–noun pairs from O*NET tasks and patents (e.g., “diagnose condition” vs. “detect cancer”) and count similarity to assign task- and then occupation-level exposure scores.
• Historical validation:
  • Robots: patents describe cleaning, moving, welding, assembling → high exposure for materials movers, equipment tenders; higher exposure associated with large declines in within-industry employment shares (moving from 25th to 75th percentile → ~9–18% decline) and wages (~8–14% decline).
  • Software: patents describe recording, storing, producing information, executing rules → exposure concentrated in middle-wage occupations; exposure associated with smaller declines (employment ~7–11%; wages ~2–6%).
• AI exposure profile:
  • Patent tasks: predicting prognosis/treatment, detecting cancer, identifying damage, detecting fraud — i.e., pattern recognition, judgment, optimization.
  • Most-exposed occupations: clinical laboratory technicians, chemical engineers, optometrists, power-plant operators, plus some inspection/quality-control production jobs.
  • Demographics: exposure rises with education (peaks near the 90th percentile), more concentrated among older workers and those with college/Master’s degrees.
• Projection caveats: projections of inequality effects assume (1) the same linear exposure → wage-change mapping as for prior technologies and (2) no correlated alternative channels (e.g., new task creation or augmentation) that systematically redistribute impacts. Webb stresses measurement, timing, and endogeneity limitations.

Data & Methods

• Text sources:
  • Patent corpus (technology-specific patents) to identify inventor-described capabilities.
  • Job-task descriptions from O*NET to represent tasks performed by occupations.
• NLP method:
  • Extract verb–noun pairs from both corpora (e.g., “diagnose condition”), compute similarity/overlap between patent verb–noun pairs and task pairs to assign a task-level automation score.
  • Aggregate task scores to occupation-level exposure metrics (and occupation–industry cells for regressions).
• Empirical strategy:
  • Case studies for software and robots: describe patented tasks, characterize demographic and occupational exposure, and estimate correlations between exposure and changes in employment and wages over 1980–2010.
  • Regressions primarily compare occupations within the same industry (occupation–industry cells) to reduce confounding from industry-level shocks; include controls for skill composition and other factors.
  • Use fitted coefficients from historical episodes to project potential impacts of AI under the maintained assumption of stable exposure → wage-change mapping.
• Theoretical framing:
  • Extends the Acemoglu–Restrepo task-based model by adding occupations and using CES aggregation at task and occupation levels; highlights ambiguous net effects of task automation on occupation demand (displacement vs. price/productivity-driven demand response), motivating empirical calibration.

Implications for AI Economics

• Measurement innovation: Webb’s patent–task overlap method provides a scalable, objective, and updatable way to map technological capabilities to occupational tasks — useful for forecasting and policy targeting without relying only on expert surveys.
• Heterogeneous impacts across technologies: different automation technologies target different tasks/skill levels (robots → manual “muscle” tasks; software → routine information tasks; AI → pattern-detection/judgment tasks). Policy responses should therefore be technology- and task-specific (education/training, social insurance, sectoral transition assistance).
• Distributional outlook for AI:
  • Unlike prior waves that disproportionately hit lower- and middle-skill jobs, AI (as observed in patents to date) is more likely to displace or transform higher-skilled, experienced occupations. This implies redistribution risks across the skill distribution are different from past automation episodes.
  • Potentially reduces 90:10 inequality under Webb’s maintained mapping, but may increase concentration at the very top (99th percentile) and could generate substantial transitional costs for displaced workers in exposed occupations.
• Caution for forecasting and policy:
  • Projections depend critically on (a) adoption speed, (b) whether AI mainly substitutes for labor vs. augments it or creates new tasks/occupations, and (c) the stability of historical exposure → outcome relationships.
  • Policymakers should combine exposure-based monitoring (patents + O*NET) with adoption metrics, firm-level case studies, and worker-level transition analyses to design targeted reskilling and social-insurance responses.
• Research use: the exposure measure can be applied in future work to track evolving AI capabilities in near real time, to study firm adoption dynamics, or to link exposure with micro-level employment transitions and welfare outcomes.