This research paper by Maxim Massenkoff and Peter McCrory introduces 'observed exposure,' a novel metric designed to quantify the actual integration of AI in the workplace. By combining theoretical task capabilities with real-world professional usage data from the Anthropic Economic Index, the authors identify a significant gap between what AI can theoretically do and its current deployment. The study finds that while digital occupations like computer programming show high exposure, there is no evidence of a systematic increase in aggregate unemployment for exposed workers since the release of ChatGPT. However, the data does suggest a 14% slowdown in the hiring of young workers (ages 22-25) in highly exposed professions. Overall, the findings indicate that AI's impact on the labor market is currently characterized more by task evolution than mass job displacement.
Key Takeaways
Actual AI coverage in professional settings remains a small fraction of its theoretical capability; for instance, only 33% of computer and math tasks show observed usage vs. a 94% theoretical potential.
Computer Programmers (74.5%) and Customer Service Representatives (70.1%) are among the most exposed occupations under the new 'observed exposure' measure.
Workers in highly exposed occupations earn approximately 47% more and are significantly more likely to hold graduate degrees compared to unexposed workers.
There has been no systematic increase in unemployment for the most AI-exposed workers since late 2022, suggesting limited displacement to date.
Hiring of young workers aged 22-25 in AI-exposed roles has decreased by roughly 14% since the release of ChatGPT.
Approximately 30% of the workforce, including roles like cooks, mechanics, and bartenders, currently has zero AI exposure due to the physical nature of their tasks.
Learning Objectives
Differentiate between theoretical AI capability and observed real-world exposure in the labor market.
Identify the demographic and professional characteristics of workers most affected by current AI integration.
Analyze the impact of AI on unemployment trends and hiring patterns for entry-level workers.
Evaluate the relationship between AI exposure and long-term occupational growth projections.
Glossary
Observed Exposure
A measurement of tasks theoretically possible with AI that are actually seeing automated, work-related usage in professional settings.
Theoretical Capability
The full range of tasks that a Large Language Model could potentially speed up or perform, regardless of current professional adoption levels.
Anthropic Economic Index
A dataset derived from real-world Claude usage patterns used to measure actual AI task coverage across different industries.
Eloundou β
A metric that scores tasks based on whether an LLM can double the speed of completion (1 = fully feasible alone, 0.5 = with tools, 0 = not feasible).
Difference-in-differences (DiD)
A statistical technique used to estimate causal effects by comparing the changes in outcomes over time between a treatment group and a control group.
Timeline
2022Baseline worker characteristics measured (August-October) and initial release of ChatGPT (late 2022).
2024Observed divergence in hiring rates where young workers became relatively less likely to be hired into AI-exposed occupations.
2025Publication of BLS employment projections (2024-2034) and collection of usage data for the Anthropic Economic Index.
2026Publication of the research paper 'Labor market impacts of AI: A new measure and early evidence'.
2034Target year for BLS employment growth projections showing slower growth in highly exposed AI roles.
Mind Map
Everything is expanded by default. Use the − buttons to collapse a branch, or the controls below.
AI Labor Impacts
Exposure Metrics
Theoretical Capability
Observed Exposure
Early Evidence
No mass unemployment spike
14% drop in youth hiring
Occupational Vulnerability
High: Computer Programmers
Zero: Physical/Service Roles
AI in the Office: Capability vs. Reality
Measuring the Gap Between Potential and Actual AI Task Coverage
computer
33%
Actual AI coverage in Computer & Math
code
74.5%
Exposure for Computer Programmers
trending_down
14%
Drop in hiring for ages 22-25
payments
47%
Higher pay for exposed workers
handyman
0%
Exposure for bartenders and cooks
The Deployment Gap
Actual AI usage remains a fraction of what is feasible due to legal, software, and human verification hurdles.
The Demographic Divide
AI-exposed workers are more likely to be female, highly educated, and in high-paying white-collar roles.
Flashcards
Tap a card to flip it.
Quiz
Frequently Asked Questions
Is AI currently causing mass unemployment?
No. Based on Current Population Survey data, researchers found no systematic increase in unemployment for highly exposed workers since late 2022. The aggregate trendlines for exposed and unexposed workers remain very similar.
Who are the most exposed workers in the current market?
The most exposed workers are typically in high-paying, white-collar roles. They are more likely to be female, older, more educated (often holding graduate degrees), and earn nearly 50% more than workers in unexposed roles.
Why is actual AI usage so much lower than what experts say is possible?
The paper highlights that theoretical potential is held back by model limitations, legal and software constraints, the requirement for human verification steps, and other organizational hurdles.
Should students entering the workforce be worried?
The study found a 14% drop in hiring for workers aged 22-25 in AI-heavy fields. This suggests that while existing experts are not being displaced, the 'entry-level' market is becoming more competitive as firms may be using AI to increase the productivity of senior staff instead of hiring juniors.
References
Labor market impacts of AI: A new measure and early evidence