AI that substitutes for workers depresses new-job creation and pay for low-skilled occupations, while AI that augments workers drives new roles and wage gains for high-skilled workers, widening pay gaps across the US labor market.

Main Finding

AI development has divergent labor-market effects depending on whether technologies automate tasks or augment worker output. Between 2015 and 2022 in the U.S., augmentation-AI development (measured via developer activity) stimulated the creation of new work and increased employment overall and raised wages in high‑skilled occupations; automation‑AI development had little measurable effect on new work or aggregate employment but reduced wages overall and, in particular, reduced new work, employment, and wages in low‑skilled occupations. These patterns imply that AI is likely contributing to rising wage inequality.

Key Points

• Distinction: The paper separates AI that substitutes tasks (automation AI) from AI that complements and expands worker output (augmentation AI). These can have opposing labor-market effects.
• New work: Augmentation AI significantly increases the emergence of new job titles (new work); automation AI shows no positive effect on new work at the aggregate level.
• Employment: Augmentation AI exposure is associated with positive employment effects overall (firms appear to hire for newly created work). Automation AI shows no significant net effect on employment overall.
• Wages: Automation AI exposure reduces average hourly wages (displacement outweighs productivity gains overall). Augmentation AI has no significant aggregate wage effect, but raises wages for high‑skill occupations.
• Heterogeneity by skill:
  • Low‑skilled occupations: Automation AI exposure leads to declines in new work, employment, and wages.
  • High‑skilled occupations: Augmentation AI exposure increases the emergence of new work and raises wages (employment effects for high‑skill are muted, possibly due to labor shortages or slow adjustment).
  • Middle‑skilled occupations: Effects lie between low and high skill results (mixed).
• Concentration: New work generation is highly concentrated, especially in computer & mathematical occupations; augmentation exposure is concentrated in STEM occupations.
• Contribution: Introduces novel measures of AI exposure (developer activity on Stack Overflow) and a new approach to identify new job titles using semantic-textual similarity.

Data & Methods

• Period and scope: U.S., 2015–2022; analysis at the occupation × industry × year level.
• Primary data sources:
  • Stack Overflow posts (AI‑related questions) to capture types of AI development by year.
  • O*NET (2015 abilities and 2015–2022 "Alternate titles") for occupational abilities and new job-title tracking.
  • 2016 Census Alphabetical Index (CAI) micro-titles to map AI applications to occupational/industry outputs.
  • Standard labor-market data for employment and hourly wages (and controls).
• Construction of exposure indices:
  • Automation‑AI index: map Stack Overflow AI questions to worker abilities (O*NET) using Semantic Textual Similarity (STS); weight by the importance of those abilities per occupation (following Felten et al. approach).
  • Augmentation‑AI index: map AI questions to micro‑titles (CAI) describing outputs; use STS to associate AI applications with specific micro-titles and aggregate to occupation–industry.
• New work measure:
  • Track changes in O*NET "Alternate titles" over time (2015–2022); identify newly added, substantively new job titles using STS and extensive cleaning to exclude mere rewordings.
• Econometric strategy:
  • Baseline: OLS regressions of outcomes (share new work, employment, wages) on automation and augmentation exposure, with extensive controls and fixed effects (including occupation×industry and year effects).
  • Causal identification: Instrumental variables (IV) using five‑year lagged AI development in countries with limited economic ties to the U.S. (argument: reduces reverse causality and common shocks; 5‑yr lag reduces anticipatory sorting).
  • Validation: Automation index correlates strongly with firm adoption measures and existing automation exposure indices (Felten et al.). Augmentation index shown to capture a distinct, interpretable dimension (task‑complementing technologies).
• Robustness: Extensive fixed effects, covariates (workforce composition, trade exposure), multiple specification checks reported.

Implications for AI Economics

• Conceptual: Demonstrates the practical importance of separating automation vs augmentation when assessing AI’s labor-market effects. Aggregated AI exposure masks opposing mechanisms and heterogeneous impacts across skill groups.
• Distributional consequences: Findings point toward widening wage inequality driven by automation pressures on low‑skill workers and augmentation benefits accruing to high‑skill workers (higher wages, new specialized roles).
• Policy relevance:
  • Labor-market policy should be targeted: protect and support low‑skill workers exposed to automation (re‑skilling, wage insurance, transition assistance).
  • Education and training: invest in upskilling that aligns with augmentation‑complementary tasks (STEM, advanced cognitive and interpersonal skills) to allow workers to capture wage gains from augmentation AI.
  • Complementary policies: encourage firm investment in worker complementarities (on‑the‑job training, job redesign) to convert productivity gains into broader employment/wage benefits.
  • Monitoring and measurement: public statistical agencies and policymakers should track both developer activity and on‑the‑job task changes (not only firm adoptions) to anticipate labor-market transitions.
• Research implications:
  • Measurement innovation: using developer‑activity data (Stack Overflow) and semantic similarity offers a promising route to observe AI development and its intended applications.
  • Need for further work on mechanisms: disentangle displacement vs productivity channels at firm and task level, longer‑run dynamics, and supply‑side responses (education, migration).
  • External validity: results are U.S.-focused and for 2015–2022—a period of rapid but still early AI diffusion—so effects may evolve as technologies mature and adoption spreads.
• Limitations to bear in mind:
  • Instrument validity depends on the assumption that AI development in selected foreign countries affects U.S. labor only via global AI technology development—not via other channels.
  • Stack Overflow reflects developer activity (development intent and capability), not direct firm adoption or on‑the‑job usage; mapping from developer posts to realized workplace change is indirect.
  • The 2015–2022 window captures early diffusion; long‑run effects (task reallocation, capital deepening) may differ.