Generative AI has not produced broad job or wage losses yet, but demand for automatable and entry-level roles has fallen while AI-complementary occupations are growing; the net long-term impact will depend on policy, reskilling, and governance.

Main Finding

Generative AI (GenAI) is already reshaping labor markets, but its effects are heterogeneous and largely task- and occupation-specific. Short-run, nationally representative data show limited aggregate wage or employment changes since the public release of ChatGPT (Nov 2022), while higher-frequency job-posting evidence reveals significant declines in demand for highly automatable and entry-level roles and concurrent growth in AI‑complementary occupations. Overall, GenAI behaves more as a task-transforming technology than a uniformly job-destroying one; policy, reskilling, and governance choices will determine whether its long-run effects widen or reduce inequality.

Key Points

• Heterogeneous impacts
  • Occupations with routine, well-specified tasks (especially entry-level positions) show the largest declines in job-posting demand.
  • Occupations that leverage human judgment, interpersonal skills, or AI supervision/augmentation show rising demand.
• Short-run aggregate outcomes
  • Nationally representative datasets (employment, earnings from household and administrative sources) show limited aggregate employment or wage declines so far.
  • Distributional shifts may be masked by slow adoption, reallocation, and measurement lags.
• High-frequency, job-market signals
  • Online job-posting data indicate rapid declines in postings for roles with high automatability and increases in postings for AI-related/complementary skill sets (prompt engineering, AI product management, data annotation, AI oversight).
• Task vs. job framing
  • GenAI substitutes for particular tasks (text generation, summarization, coding scaffolding) rather than entire occupations in many cases; this creates mixed complementarities within occupations.
• Emerging dynamics
  • Firms reorganize job bundles and create new task compositions; some workers upskill into higher-value tasks, while others face displacement or stagnating wages.
• Policy sensitivity
  • Outcomes depend heavily on retraining availability, labor-market mobility, wage-setting institutions, and AI governance (deployment limits, safety, transparency).

Data & Methods

• Data sources used in the paper
  • Population-level datasets: nationally representative household surveys and administrative records (employment, earnings) to assess aggregated labor-market outcomes.
  • Online job-posting platforms: high-frequency indicators of labor demand by occupation, task keywords, and required skills.
  • Systematic literature review/meta-analysis: synthesizing emerging micro- and macro-level empirical studies on GenAI impacts.
• Empirical approaches commonly applied
  • Event-study / difference-in-differences exploiting the public release of ChatGPT (Nov 2022) or other sudden adoption episodes as quasi-experiments.
  • Occupational exposure measures: construction of AI-exposure indices using task content and model capabilities to classify occupations by automatability and complementarity.
  • Text analysis / skill tagging: natural language processing on job ads to detect demand shifts for specific skills (e.g., prompt engineering, AI tools).
  • Heterogeneity analysis by worker characteristics (education, experience), firm size, sector, and geography.
• Limitations noted
  • Short evaluation window since public ChatGPT release; long-run effects remain uncertain.
  • Job-posting data may not fully capture filled positions, informal hiring, or internal reassignments.
  • Measurement error in AI exposure indices and—in some administrative data—delayed reporting.
  • Selection and adoption heterogeneity across firms complicate causal attribution.

Implications for AI Economics

• Theory and modeling
  • Models should emphasize tasks rather than occupations, incorporate capital-skill complementarity, and include reallocation frictions, learning-by-doing, and endogenous skill upgrading.
  • Balance short-run partial-equilibrium evidence with potential long-run general-equilibrium effects (productivity gains, price changes, demand shifts).
• Empirical priorities
  • Build and maintain real-time, linked employer–employee datasets and richer task/skill taxonomies to monitor distributional impacts.
  • Focus on long-run wage dynamics, within-occupation task reallocation, firm-level adoption heterogeneity, and cross-country comparisons.
• Policy recommendations
  • Invest in scalable reskilling and lifelong learning targeted at upgrading workers toward AI-complementary tasks (supervision, interpretation, complex judgment).
  • Strengthen active labor-market policies (job-search assistance, portable credentials) and reduce frictions to occupational mobility.
  • Design AI governance that mitigates concentrated displacement risks: transparency, auditing of deployed systems, sector-specific deployment limits where necessary.
  • Consider redistribution and safety-net measures (income support, wage insurance) to smooth transition costs and limit inequality amplification.
• For stakeholders (firms, educators, regulators)
  • Firms: adopt change management emphasizing worker augmentation, internal retraining, and redesign of job bundles to capture productivity gains broadly.
  • Educators/training providers: update curricula to combine domain expertise with AI literacy, human-AI interaction skills, and higher-order problem solving.
  • Regulators: monitor labor-market indicators, encourage disclosure on AI adoption, and coordinate policies that preserve inclusive gains from GenAI.