Abstract
Since the public release of ChatGPT in late 2022, the economic discourse has been dominated by a tension between potential productivity gains and the threat of mass technological unemployment. This paper provides empirical evidence on this debate using a Synthetic Difference-in-Differences (SDiD) approach. We find that the "ChatGPT shock" has resulted in a significant increase in weekly earnings (+$89) for workers in occupations highly exposed to Large Language Models (LLMs), while having no discernible impact on unemployment rates. These findings suggest that, in the short term, LLMs are functioning as productivity-enhancing complements rather than job-replacing substitutes.
Introduction
The narrative surrounding Artificial Intelligence (AI) is often polarized. On one hand, proponents argue that AI will usher in an era of unprecedented productivity; on the other, critics fear a wave of automation that renders human labor obsolete. While early research focused on the potential for automation based on task descriptions, empirical data on realized labor market outcomes has been scarce.
This study addresses the "substitution vs. complementarity" debate. In economic theory, if LLMs are substitutes, they should decrease wages and employment as machines replace human workers. Conversely, if they are complements, they should increase worker productivity, shifting the labor demand curve outward and raising wages. By analyzing data from the post-ChatGPT era, this paper provides some of the first large-scale evidence to determine which force is currently dominant in the labor market.
Methodology: Synthetic Difference-in-Differences
To isolate the causal impact of LLM exposure, we employ a Synthetic Difference-in-Differences (SDiD) methodology.
Standard Difference-in-Differences (DiD) relies on the assumption that "treated" (AI-exposed) and "control" (non-exposed) groups would follow parallel trends in the absence of the treatment. However, the post-COVID economy is characterized by volatile shifts that make this assumption difficult to satisfy. SDiD addresses this by constructing a "synthetic" control group, reweighting less-exposed occupations to match the pre-2022 trends of highly exposed occupations. This approach ensures a more robust causal inference by accounting for pre-existing differences between occupational categories.
Defining Exposure
Rather than relying on subjective expert assessments, we utilize a novel "prompt-based" measure of exposure. By mapping occupational tasks to millions of real-world user prompts sent to LLMs (specifically Anthropic’s Claude), we identify which jobs are truly susceptible to LLM influence. This method captures the actual utility of the technology in a professional context.
Data Sources
The study integrates several high-fidelity datasets to track labor market shifts:
- Current Population Survey (CPS): We utilize Outgoing Rotation Group (ORG) data from January 2010 through August 2025, providing a comprehensive time series of earnings and employment status.
- O*NET (Occupational Information Network): This allows for the decomposition of occupations into specific constituent tasks.
- Handa et al. (2025) Exposure Data: This dataset links O*NET tasks to millions of LLM prompts, classifying tasks as "automative," "augmentative," or neither.
Key Findings
1. The Earnings Boost
The most significant finding is a clear "wage premium" for AI-exposed roles. Occupations with high LLM exposure saw an average increase of $89 in weekly earnings (adjusted to 2010 prices) compared to the synthetic control group. This earnings gap began widening significantly in 2024 and 2025, suggesting a lag between the initial technological shock and the manifestation of productivity gains in worker paychecks.
2. Employment Stability
Despite fears of displacement, the impact on unemployment has been negligible. We observed a statistically insignificant change of only 0.2 percentage points in unemployment rates for exposed occupations. This suggests that firms are currently using LLMs to make existing workers more efficient rather than replacing them with automated systems.
3. Inelastic Labor Supply
A central question is why these higher wages have not yet triggered a massive influx of workers into exposed fields, which would theoretically drive wages back down. The data suggests that labor supply remains inelastic in the short term due to "frictions"—specifically the time and effort required for workers to acquire the new skills necessary to leverage LLMs effectively. This allows incumbent workers in these fields to capture the productivity gains as higher earnings.
Discussion and Mechanisms
The divergence between rising wages and stable employment points toward a complementarity story. If AI were primarily automating jobs, we would expect to see rising unemployment and downward pressure on wages. Instead, the "Intent-to-Treat" analysis—measuring potential exposure rather than realized adoption—reveals that the mere availability and utility of these tools are driving value. By focusing on potential exposure, we avoid "endogeneity" risks, ensuring that the results reflect the technology's impact rather than simply identifying firms that were already high-performing.
Limitations and Future Outlook
While the current data is optimistic for workers, several caveats remain:
- Short-Term vs. Long-Term: These results are explicitly short-term. As technology evolves or as firms undergo structural reorganizations to favor automation, LLMs could transition from complements to substitutes.
- Entry-Level Vulnerability: While average effects are positive, other research suggests that entry-level workers may face greater displacement pressure than mid-career professionals who possess the domain expertise to guide LLM outputs.
- Heterogeneity of Impact: The positive average hides underlying variance. A subset of occupations did experience earnings declines, indicating that the transition creates both winners and losers depending on the specific nature of the tasks involved (e.g., creative writing vs. routine data entry).
Conclusion
In the immediate wake of the ChatGPT release, the labor market has responded not with mass unemployment, but with a productivity-driven wage increase for exposed occupations. For now, the "ChatGPT Dividend" is being captured by workers, suggesting that LLMs are currently serving as a powerful tool for human augmentation.
Summary updated