LLM Impact on Labor Markets

• LLM labor market impact is defined by quantifiable exposure indices derived from task-level rubrics and expert assessments.
• Empirical studies reveal that LLMs affect high-wage, text-driven sectors, leading to shifts in earnings, job demand, and skill transitions.
• Forecasting methods employing LLMs, including time series and quasi-experimental designs, provide actionable insights into occupational and industry adjustments.

LLMs are general-purpose, foundation-model architectures capable of producing, interpreting, and reasoning over human language and code at scale. Their rapid advances have catalyzed new frameworks for assessing labor market exposure, forecasting occupational demand shifts, and understanding displacement and augmentation dynamics across diverse sectors. This article synthesizes the state of empirical measurement, theoretical modeling, prediction methodologies, and policy implications for the labor market impact potential of LLMs, as established in the recent literature.

1. Frameworks for Quantifying LLM Labor Market Exposure

The canonical approach to quantifying occupational and industrial exposure to LLMs relies on task-level rubrics informed by explicit model capability assessments and expert judgment. Major contributions include exposure indices derived from O*NET tasks and abilities, weighted by the relatedness to LLM capabilities, and aggregation to occupation and industry scales using employment shares (Eloundou et al., 2023, Felten et al., 2023).

Eloundou et al. define a three-tier rubric for each Detailed Work Activity (DWA) or task: E0 (no exposure), E1 (direct speedup by ≥50% via LLM alone), E2 (≥50% speedup with LLM-powered software), and E3 (requiring multimodal LLMs) (Eloundou et al., 2023). Weighted task scores are aggregated to occupational exposure scores β_i using:

$$\beta_i = \sum_j w_{ij} \cdot (I_{ij}^{E1} + 0.5 I_{ij}^{E2})$$

where w_{ij} is the task weight (core tasks doubly weighted). Empirically, 80% of workers are in occupations with β_i ≥ 0.10, and 19% with β_i ≥ 0.50; mean β_i at the task level is approximately 31% (human raters), increasing to 56% for the upper bound with all LLM-software exposure.

Industry and geographic exposure are aggregated by weighting occupational exposure scores with local or sectoral employment shares (Felten et al., 2023). In China, similar task-based indices built using both human and LLM classifiers show mean LLM-based exposure correlating positively (ρ≈0.4) with wage and experience premiums, deviating from the routine-task focus of past automation (Chen et al., 2023).

2. Heterogeneous Sectoral and Occupational Effects

LLM exposure is concentrated in higher-wage, text-driven, and cognitively intensive occupations and industries. In the U.S., occupations such as telemarketers (AIOE_k = 1.926), post-secondary language and history teachers (AIOE_k = 1.81–1.85), and law teachers dominate the exposure distribution (Felten et al., 2023). The most exposed sectors by AIIE_m include legal services, financial investments, and insurance, reflecting the prevalence of textual analysis, report generation, and document processing.

Cross-nationally, Chinese labor market studies find that Education, Culture & Arts, and Healthcare have the highest industry-level exposure r_i, while Manufacturing, Agriculture, Mining, and Construction remain the least exposed (Chen et al., 2023). Notably, these exposure scores positively correlate with both wage and experience returns, indicating threats to both routine and nonroutine high-income roles, breaking the classic routinization paradigm.

3. Dynamic Labor Market Adjustments: Displacement, Augmentation, and Skill Transition

The labor market effects of LLM deployment are multifaceted, operating through intertwined channels of substitution, complementarity, and task creation. Liu et al. demonstrate that the release of ChatGPT caused a statistically significant contraction in gig-level transaction volumes for AI-exposed categories (–3.1% at the gig level, –3.9% at the freelancer level), with the displacement effect amplified for low-quality and high-volume sellers (Liu et al., 2023). High-quality freelancers exhibited substantially milder declines (–2.68%), while those embracing AI by offering LLM-centric gigs saw positive effects (+7.81%).

Quasi-experimental evidence from online labor platforms reveals occupation-specific heterogeneity. For example, LLM deployment led to pronounced displacement in translation (jobs –7.4%, earnings –30.2%) and writing (jobs –4.5%), but strong productivity-driven increases in web development (jobs +7.3%, earnings +60%) and machine learning (jobs +5.9%, earnings +38.2%) (Qiao et al., 2023). A Cournot competition model formalizes phases: a “honeymoon” of productivity gain below an inflection point in LLM capability, and “substitution”—net harm as AI capability passes the point that further advancements sharply reduce human demand.

Skill transitions are a salient consequence. By lowering barriers to entry in programming and other submarkets, LLMs induce incumbents and high-skilled freelancers to reallocate supply toward AI-augmented jobs, intensifying competition (Liu et al., 2023).

4. Forecasting and Predictive Methodologies Leveraging LLMs

A new genre of research explores the direct use of LLMs as forecasting engines for labor demand. The benchmark developed in "How AI Forecasts AI Jobs" formalizes labor forecasting as a time series extrapolation problem, using both high-frequency sectoral job postings and annual AI-occupational growth indices (Osborn et al., 27 Oct 2025). LLMs are conditioned on historic numeric inputs, with output structured in JSON for automated evaluation. Forecast accuracy is measured by MSE, RMSE, and MAE.

Prompt engineering—task-scaffolded, relative, event reasoning, persona-driven—substantially modulates forecast stability and accuracy. Persona prompts grounded in HR or research roles yield more reliable outputs. For long-horizon forecasts (3–5 years), event-based causal prompts outperform pure extrapolation, integrating exogenous shocks such as AI breakthroughs. On short horizons, moving average statistical baselines remain competitive, with only marginal LLM outperformances. Model scale is less decisive than prompt architecture: smaller, prompt-optimized models (e.g. GPT-4o-mini) can outperform larger models (LLaMA-70B) (Osborn et al., 27 Oct 2025).

Online discussion indices (Reddit/news) of generative AI serve as empirically validated leading indicators for labor market shifts, forecasting movements in job openings, hiring rates, tenure, and unemployment with 1–7 month lead times, especially in knowledge-intensive fields (Cao et al., 20 Nov 2025).

5. Labor Market Mechanisms: Wages, Earnings, and Unemployment

Causal designs (synthetic difference-in-differences, event-study) show that LLM exposure initially increases earnings in high-exposure occupations without raising unemployment (Chen et al., 19 Sep 2025). After ChatGPT’s public launch, average weekly wages in exposed occupations rose by ≈\$89, with no significant movement in unemployment rates—suggesting that the first-order channel is productivity enhancement, not displacement. These findings are robust to varying exposure measures (augmentation, automation, continuous vs. binary treatment) and placebo specifications.

In online labor markets, LLMs disrupt the allocation of work by collapsing the cost of written communication, undermining costly signaling mechanisms that previously allowed higher-ability workers to differentiate themselves (Galdin et al., 11 Nov 2025). Structural equilibrium analysis shows that after LLM-driven loss of written application signaling, high-ability workers are hired 19% less often, and low-ability workers are hired 14% more often, reducing overall market meritocracy even if fill rates and wage dispersion compress only moderately.

6. Skill Requirement Evolution, New Occupations, and Recruitment

LLM adoption is rapidly shifting skill requirements. In China, 28% of occupations in 2023 already required ChatGPT-related skills; projections indicate that an additional 45% will require such skills within 2–3 years, with proficiency levels differentiated by industry. Salaries for ChatGPT-enabled roles register 8–15% premia over occupation baselines (Chen et al., 2023). The most rapid uptake occurs in technology, product, operations, marketing, and design; manufacturing, education, and healthcare integrate ChatGPT skills more gradually. New roles (prompt engineer, LLM trainer, AI-process optimizer) are emerging across labor markets.

In recruitment, LLMs can produce implicit weighted “fit utilities” for candidate–job pairings, emphasizing experience, skills match, and reputation, with minimal mean discrimination by gender/ethnicity but nuanced intersectional effects. Automated LLM shortlisting improves screening speed and match quality but poses risks of amplified intersectional biases, which may cumulate over multi-stage recruitments. Regulatory frameworks are recommended, including baseline audits for group-wise disparity ratios and correction mechanisms for undesired interaction effects (Hoffmann et al., 16 Jan 2026).

7. Modeling Labor Market Shocks and Systemic Implications

Advanced embedding-based frameworks, such as Labor Space, produce high-dimensional vector representations of occupations, skills, industries, and firms, fine-tuned on taxonomic and relational information (Kim et al., 2023). This enables mapping of LLM proximity across the labor market and forecasting the propagation of LLM-induced shocks by the diffusion of exposure via network adjacency matrices. High cosine similarity to LLM embeddings strongly predicts exposure rankings and aligns with external occupational and industry exposure indices.

LLM penetration produces ripple effects, shifting skill demand distributions, polarizing wage structures, and altering firm strategic positioning. These effects are not confined to routine jobs; LLMs threaten high-wage, high-experience nonroutine roles and may compress experience premiums (Chen et al., 2023). Policy responses must balance productivity gains from LLM adoption with disruption costs, setting sector-specific thresholds for socially optimal diffusion.

---

References:

• "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of LLMs" (Eloundou et al., 2023)
• "How will Language Modelers like ChatGPT Affect Occupations and Industries?" (Felten et al., 2023)
• "Generate the Future of Work through AI: Empirical Evidence from Online Labor Markets" (Liu et al., 2023)
• "The (Short-Term) Effects of LLMs on Unemployment and Earnings" (Chen et al., 19 Sep 2025)
• "How AI Forecasts AI Jobs: Benchmarking LLM Predictions of Labor Market Changes" (Osborn et al., 27 Oct 2025)
• "Evaluating LLM Behavior in Hiring: Implicit Weights, Fairness Across Groups, and Alignment with Human Preferences" (Hoffmann et al., 16 Jan 2026)
• "AI and Jobs: Has the Inflection Point Arrived? Evidence from an Online Labor Platform" (Qiao et al., 2023)
• "LLMs at Work in China's Labor Market" (Chen et al., 2023)
• "Labor Space: A Unifying Representation of the Labor Market via LLMs" (Kim et al., 2023)
• "Can Online GenAI Discussion Serve as Bellwether for Labor Market Shifts?" (Cao et al., 20 Nov 2025)
• "Making Talk Cheap: Generative AI and Labor Market Signaling" (Galdin et al., 11 Nov 2025)
• "The Future of ChatGPT-enabled Labor Market: A Preliminary Study in China" (Chen et al., 2023)
• "The Impact of Large Language Multi-Modal Models on the Future of Job Market" (Singh, 2023)