Displacement or Complementarity? The Labor Market Impact of Generative AI

The paper investigates whether generative AI primarily displaces workers by automating cognitive tasks or complements human labor by augmenting productivity, thereby increasing labor demand. The context is the rapid diffusion of generative AI technologies (e.g., ChatGPT since late 2022) that can perform text generation, coding, summarization, and decision support—capabilities historically requiring human expertise. The study formulates the central hypothesis that generative AI reduces job postings in automation-prone occupations and increases postings in augmentation-prone occupations. It addresses a gap in prior work that largely relies on task-exposure models without directly examining labor market outcomes. Using indices of automation and augmentation potential at the occupation level and a near-universe of U.S. job postings, the paper evaluates changes in labor demand and evolving skill requirements post-introduction of generative AI.

The paper builds on theories of task-based technological change and automation’s effects on labor demand (Acemoglu & Restrepo, 2018, 2019, 2020) and skill-biased technological change (Autor et al., 2003, 2006). Prior empirical work has used AI exposure indices (Felten et al., 2018, 2023; Brynjolfsson & Mitchell, 2017; Brynjolfsson et al., 2018; Webb, 2019) to study employment, wages, and productivity but offers limited insight into AI’s potential to augment human labor. Recent studies assess generative AI’s productivity impacts in specific settings (e.g., customer support, online platforms; Brynjolfsson et al., 2023; Hui et al., 2024; Liu et al., 2023; Noy & Zhang, 2023) and task exposure (Eloundou et al., 2024), yet they do not comprehensively analyze economy-wide labor demand and skills. This study contributes by developing both automation and augmentation indices and empirically testing displacement versus complementarity effects using job postings and skill requirements across U.S. occupations.

Data sources: ONET v25.1 task descriptions covering 19,265 tasks across 923 occupations and Lightcast U.S. online job postings with posting dates, titles, and required skills (2019–June 2024). Task exposure scoring: Using GPT-4o, each task is classified into E0 (no exposure), E1 (direct LLM exposure), E2 (exposure via LLM-powered applications), or E3 (exposure given image capabilities). Occupation automation score: weighted sum over tasks with weights 1 for E1, 0.5 for E2, and 0 for E0 and E3, further weighted by ONET task importance, following Eloundou et al. (2024). Augmentation score: computed via a Herfindahl-based measure that is highest when an occupation mixes exposed and non-exposed tasks. Specifically, AI-exposed tasks are E1, E2, E3; non-exposed are E0. The augmentation score equals 1 minus the sum of squared shares of exposed and non-exposed tasks (each share weighted by task importance). Skills classification: Using Lightcast’s 33,620 unique skills, ChatGPT classifies each skill into S0 (irrelevant to GenAI), S1 (GenAI-relevant), S2 (complemented by GenAI), or S3 (substituted by GenAI). For analysis, S1–S3 are grouped as AI-exposed skills, S0 as non-exposed. New skills: Baseline required skills per firm-occupation are set from 2015–2019; from 2020 onward, any skill not in the baseline appearing in a firm-occupation is labeled new and added to the baseline for future tracking. Aggregation: For each firm and quarter beginning in 2019, outcomes include counts of job postings, total required skills, AI-exposed skills, and new skills, computed separately for quartiles of automation and augmentation scores. Research design: The introduction of ChatGPT (November 2022) is treated as an exogenous shock. Treatment groups are the top quartiles of automation or augmentation; controls are quartiles 3 and 4 (excluding quartile 2 to reduce spillovers). Primary estimation uses Synthetic Difference-in-Differences (SDID; Arkhangelsky et al., 2021), constructing synthetic controls that match pre-treatment trajectories and producing event study plots to assess dynamic effects and pre-trends. An alternative specification employs Poisson DiD for count data, weighting controls by SDID-derived synthetic weights to balance pre-treatment trends. Pre-period spans 2019–Q3 2022; post-period spans Q4 2022–Q2 2024. Robustness includes block-bootstrap for SDID standard errors and clustering at the firm level for Poisson models.

• Labor demand: Occupations in the top automation quartile experienced an average decline of 95.4 job postings per firm per quarter post-ChatGPT (SDID), corresponding to a 17% decline in postings (Poisson DiD estimate -0.169). Augmentation-prone occupations saw an average increase of 79.8 postings per firm per quarter (SDID), equating to a 22.5% rise (Poisson DiD estimate 0.225). Event studies show parallel pre-trends and persistent post-treatment divergence. - Skills in automation-prone occupations: Significant declines post-ChatGPT in total required skills (-297.7, SDID; -23.8%, Poisson), AI-exposed skills (-221.3, SDID; -24.0%, Poisson), and new skills (-30.6, SDID; -38.1%, Poisson). - Skills in augmentation-prone occupations: Significant increases post-ChatGPT in total required skills (+67.3, SDID; +14.8%, Poisson), AI-exposed skills (+38.0, SDID; +15.1%, Poisson), and new skills (+3.63, SDID; +16.7%, Poisson). - Exposure distributions: Tasks—E0 55.1%, E1 28.3%, E2 16.2%, E3 0.4%. Skills—S0 48.9%, S1 12.2%, S2 36.5%, S3 2.45%. - Index variability: Automation index mean 0.316 (SD 0.157); augmentation index mean 0.298 (SD 0.197). Top automation occupations include correspondence clerks, interpreters/translators, court and license clerks; top augmentation occupations include clinical neuropsychologists, medical dosimetrists, and first-line supervisors of police and detectives.

The findings directly address the displacement-versus-complementarity question. Generative AI reduces demand for occupations composed largely of automatable cognitive tasks, evidencing displacement. Simultaneously, it raises demand and broadens skill requirements in occupations where AI can augment workers by partially automating tasks and enabling focus on higher-value activities. Skill dynamics mirror these mechanisms: automation simplifies roles and decreases both existing and emerging skill needs, whereas augmentation increases AI-exposed, total, and new skills, indicating upskilling and the emergence of human-AI collaboration capabilities. These results imply heterogeneous labor market effects and suggest that generative AI operates as both a capital-augmenting and labor-augmenting technology depending on occupational task composition. Policy and managerial implications include targeted support for workers in highly automatable roles, investment in AI literacy and complementary skills, and organizational redesign to harness productivity gains from human-AI collaboration.

Generative AI exerts dual impacts on the labor market. In automation-prone occupations, it simplifies tasks and reduces the demand for specialized skills and postings. In augmentation-prone occupations, it enhances productivity and increases demand for broader, more advanced skill sets and hiring. Recognizing these heterogeneous effects is essential for policymakers and firms to craft strategies that mitigate displacement risks and foster skill development for effective human-AI collaboration as AI adoption scales.

The analysis captures short-term impacts during early generative AI adoption; long-term effects remain uncertain. Job postings may imperfectly proxy labor demand due to phenomena like ghost postings. Results reflect the U.S. labor market and may not generalize to regions with different adoption rates or labor structures. Measurement relies on GPT-based task and skill classifications and O*NET importance weights, which, despite validation and robustness, may introduce classification or weighting biases.