The (Short-Term) Effects of Large Language Models on Unemployment and Earnings (2509.15510v1)

Abstract: LLMs have spread rapidly since the release of ChatGPT in late 2022, accompanied by claims of major productivity gains but also concerns about job displacement. This paper examines the short-run labor market effects of LLM adoption by comparing earnings and unemployment across occupations with differing levels of exposure to these technologies. Using a Synthetic Difference in Differences approach, we estimate the impact of LLM exposure on earnings and unemployment. Our findings show that workers in highly exposed occupations experienced earnings increases following ChatGPT's introduction, while unemployment rates remained unchanged. These results suggest that initial labor market adjustments to LLMs operate primarily through earnings rather than worker reallocation.

• The paper employs a Synthetic Difference-in-Differences method to identify the short-term causal effects of LLM diffusion on labor outcomes.
• The analysis finds that high LLM exposure is associated with an average weekly wage increase of $89 while unemployment remains largely unchanged.
• The study differentiates between augmentation and automation effects, indicating that LLMs enhance productivity without displacing workers in the short run.

Short-Term Labor Market Effects of LLMs: Evidence from Synthetic Difference-in-Differences

Introduction

This paper investigates the short-term labor market impacts of LLMs, focusing on the period following the public release of ChatGPT in late 2022. The analysis centers on two primary outcomes: unemployment and earnings, stratified by occupational exposure to LLMs. The authors employ a Synthetic Difference-in-Differences (SDiD) methodology to estimate causal effects, leveraging quasi-experimental variation in LLM exposure across occupations. The paper provides empirical evidence that, in the immediate aftermath of LLM diffusion, wage adjustments rather than employment reallocation constitute the primary margin of labor market response.

Measurement of LLM Exposure

Occupational exposure to LLMs is operationalized using a prompt-based mapping between O*NET tasks and LLM capabilities, as developed in prior work. Each occupation's exposure is quantified as the share of its constituent tasks that are susceptible to LLM augmentation or automation, based on observed prompt distributions from Anthropic Claude. This approach captures the technological frontier of LLM applicability rather than realized adoption, mitigating endogeneity concerns but potentially overstating short-run effects if actual adoption lags technical feasibility.

Figure 2: Distribution of LLM exposure across occupations, with the median marked by a red line.

The distribution of exposure is highly skewed, with a mode at zero and a long right tail. The most exposed occupations are concentrated in writing, coding, and information technology domains.

Data and Empirical Strategy

The analysis utilizes monthly microdata from the Current Population Survey (CPS) spanning January 2010 to August 2025. Earnings are measured using the Outgoing Rotation Group (ORG) sample, with all values deflated to January 2010 dollars. Unemployment is defined as the share of the labor force actively seeking work within each occupation.

The empirical strategy begins with a canonical two-way fixed effects Difference-in-Differences (DiD) design, treating the introduction of ChatGPT as a discrete shock. However, pre-treatment trend violations—particularly for unemployment—necessitate the use of SDiD, which reweights control occupations and pre-treatment periods to better match the treated group’s pre-trends, thereby addressing latent factor confounding.

Descriptive Trends

Unemployment rates are consistently lower in occupations with higher LLM exposure, both before and after the introduction of ChatGPT. The only major deviation is the transitory spike during the COVID-19 pandemic. Notably, there is no evidence of a post-LLM increase in unemployment for highly exposed occupations.

Figure 3: Unemployment by quartile of LLM exposure, showing stability across groups except for the COVID-19 period.

Earnings trajectories reveal that high-exposure occupations not only have higher baseline wages but also experience a pronounced post-2022 acceleration in wage growth. Between December 2022 and August 2025, the top two quartiles of exposure saw an average weekly earnings increase of $105, compared to$60 in the lower half.

Difference-in-Differences and Event Study Results

Event paper analyses confirm the presence of pre-trend violations for unemployment, undermining the parallel trends assumption required for causal inference in standard DiD. For earnings, pre-trends are more stable, but some deviations persist.

Figure 4: Event paper of unemployment, showing no significant post-treatment divergence between high- and low-exposure occupations.

Synthetic Difference-in-Differences Estimates

The SDiD estimator, applied at the occupation level, yields the following key results:

• Unemployment: The distribution of estimated treatment effects is tightly centered around zero, with a mean effect of 0.2 percentage points. There is no systematic evidence that LLM exposure increased unemployment in the short run.

Figure 1: Distribution of SDiD unemployment effects, centered near zero.

• Earnings: The distribution of estimated effects is predominantly positive, with an average increase of $89 per week for highly exposed occupations. Some heterogeneity is observed, with a minority of occupations experiencing negative effects, but the modal response is a wage gain.

  Figure 5: Distribution of SDiD weekly earnings effects, with most mass above zero.

These results are robust to alternative exposure measures (augmentation vs. automation) and to continuous rather than binary treatment definitions.

Heterogeneity by Exposure Type

Disaggregating exposure into augmentation and automation reveals similar qualitative patterns. Augmentation exposure is associated with positive wage effects and negligible unemployment changes, while automation exposure does not yield significant unemployment increases in the short run.

Figure 6: Unemployment by quartile of augmentation exposure, showing no post-treatment increase.

Figure 7: Unemployment by quartile of automation exposure, also showing stability.

Theoretical and Practical Implications

The findings support a complementarity interpretation: in the short run, LLMs act as productivity-enhancing tools that increase labor demand in exposed occupations, with wage adjustments absorbing the shock due to inelastic short-run labor supply. The absence of unemployment effects suggests that, at least initially, LLMs augment rather than substitute for human labor in most affected domains.

This stands in contrast to some prior studies reporting negative employment effects for entry-level or automation-prone roles, highlighting the importance of exposure measurement and temporal scope. The results also imply that realized adoption and task reorganization may lag technical feasibility, and that labor market frictions (e.g., occupational mobility constraints) play a significant role in mediating the impact of technological shocks.

Limitations and Future Directions

The analysis is limited to short-term effects and does not capture potential medium- or long-term displacement as LLM adoption deepens and as complementary investments (e.g., workflow redesign, capital reallocation) materialize. The exposure metric, while exogenous, may overstate realized adoption. Further, the SDiD approach, while robust to latent factor confounding, cannot address all sources of endogeneity, such as unobserved shocks correlated with both LLM exposure and labor market outcomes.

Future research should track longitudinal adoption patterns, explore heterogeneous effects by demographic group (e.g., age, education), and examine downstream impacts on job quality, task content, and career trajectories. The interplay between LLM-driven augmentation and automation will likely evolve as models acquire new capabilities (e.g., multimodal reasoning), necessitating dynamic exposure measurement.

Conclusion

This paper provides evidence that, in the immediate aftermath of LLM diffusion, the primary labor market response is an increase in earnings for highly exposed occupations, with no corresponding rise in unemployment. These results are consistent with short-run productivity gains and inelastic labor supply, rather than displacement or reallocation. The findings underscore the importance of distinguishing between augmentation and automation effects and highlight the need for ongoing monitoring as LLM capabilities and adoption patterns evolve.