Labor Market Impacts of AI: A New Measure and Early Evidence
Key Findings
- A new metric called "observed exposure" combines theoretical LLM capability with real-world usage patterns, emphasizing automated and work-related applications
- Current AI adoption remains significantly below theoretical potential across most sectors
- Positions with higher exposure show weaker employment projections through 2034 according to BLS data
- Workers in highly exposed fields tend to be older, female, more educated, and earn substantially more
- No significant unemployment increase detected among exposed workers since late 2022, though hiring of young workers in exposed fields shows tentative decline
Introduction
The paper establishes a framework for measuring AI's labor market effects before major disruptions become apparent. Previous attempts to predict job displacement—such as offshorability studies and robot impact research—have produced inconsistent results, suggesting the need for careful measurement before effects fully materialize.
The authors note that "impacts of AI, however, might be less like COVID and more like the internet or trade with China," making early detection challenging through aggregate unemployment data alone.
Measuring Exposure
The analysis integrates three data sources:
- O*NET database - Catalogs approximately 800 occupational categories with associated tasks
- Anthropic usage data - Demonstrates actual Claude implementation in professional contexts
- Eloundou et al. (2023) assessments - Evaluate theoretical LLM capability to accelerate tasks by 50% or more
The theoretical capability metric (beta) scores tasks as: 1 (LLM alone), 0.5 (requires additional tools), or 0 (not feasible).
Observed Exposure Methodology
"Observed exposure" quantifies which theoretically feasible tasks actually receive automated professional use. This bridges the gap between capability and implementation, accounting for legal constraints, software requirements, and verification procedures that slow adoption.
The measure weights jobs higher when:
- Tasks show theoretical AI feasibility
- Work-related usage appears in Claude traffic
- Automated rather than augmentative implementations dominate
- AI-affected tasks comprise larger portions of the role
Figure 2 illustrates the substantial gap: "Claude currently covers just 33% of all tasks in the Computer & Math category," despite 94% theoretical feasibility.
Most Exposed Occupations
Computer Programmers rank highest with 75% coverage, followed by Customer Service Representatives and Data Entry Keyers at 67%. Approximately 30% of workers show zero detectable exposure, including cooks, mechanics, lifeguards, and bartenders.
Exposure and Employment Projections
Analysis against BLS 2024-2034 projections reveals that "for every 10 percentage point increase in coverage, the BLS's growth projection drops by 0.6 percentage points." This modest correlation provides some validation, though the relationship is subtle. Notably, theoretical capability measures alone show no such correlation.
Worker Characteristics
Highly exposed workers (pre-ChatGPT, August-October 2022) demonstrate distinct demographics:
- 16 percentage points more likely to be female
- 11 percentage points more likely to be white
- Nearly double the likelihood of being Asian
- 47% higher average earnings
- Graduate degree holders represent 17.4% (versus 4.5% in unexposed groups)
Unemployment Analysis
Comparing top-quartile exposure workers against unexposed groups reveals:
No systematic unemployment increase detected post-ChatGPT release. The difference-in-differences analysis shows the gap between groups remains statistically indistinguishable from zero.
The framework could detect differential unemployment increases around 1 percentage point. A "Great Recession for white-collar workers" scenario—doubling top-quartile unemployment from 3% to 6%—would be visible in current data.
Young Worker Hiring Trends
While overall unemployment shows no clear pattern, young workers (22-25) entering exposed occupations demonstrate a tentative signal:
"Job finding rates at the less exposed occupations remain stable at 2% per month, while entry into the most exposed jobs decreases by about half a percentage point."
This 14% relative decline is barely statistically significant and admits alternative interpretations—job stayers, career changers, or measurement error in survey transitions.
Discussion
The research establishes an updatable framework for tracking AI's employment effects before major disruptions materialize. Currently, "no impact on unemployment rates for workers in the most exposed occupations" appears evident, though hiring patterns warrant continued monitoring.
Future refinements should incorporate updated usage data, recalibrate capability assessments as models advance, and examine recent graduates' labor market trajectories in exposed credential areas.
Citation: Massenkoff, M., & McCrory, P. (2026, March 5). Labor market impacts of AI: A new measure and early evidence. Anthropic.