Key findings

Researchers introduce a new measure of AI displacement risk called observed exposure, which combines theoretical LLM capability and real-world usage data. They find that AI is far from reaching its theoretical capability, with actual coverage remaining a fraction of what's feasible. Occupations with higher observed exposure are projected to grow less through 2034, according to the BLS. Workers in the most exposed professions tend to be older, female, more educated, and higher-paid. So far, there's no systematic increase in unemployment for highly exposed workers since late 2022, but there's suggestive evidence that hiring of younger workers has slowed in exposed occupations.

Introduction

The thing is, measuring AI's labor market impacts is tricky. Past approaches have had limited success. For example, a prominent attempt to measure job offshorability identified roughly a quarter of US jobs as vulnerable, but a decade on, most of those jobs maintained healthy employment growth. The government's own occupational growth forecasts have added little predictive value beyond linear extrapolation of past trends. Even in hindsight, the impact of major economic disruptions on the labor market is often unclear. So, researchers are presenting a new framework for understanding AI's labor market impacts and testing it against early data.

Counterfactuals

Causal inference is easier when the effects are large and sudden, like during the COVID-19 pandemic. However, AI's impacts might be more subtle, like the internet or trade with China. To isolate AI's effect, researchers compare outcomes between more or less AI-exposed workers, firms, or industries. They define exposure at the task level, considering whether AI can perform specific tasks. The researchers follow this task-based approach, incorporating measures of theoretical AI capability and real-world usage.

Measuring exposure

The researchers combine data from three sources: the O*NET database, their own usage data from the Anthropic Economic Index, and task-level exposure estimates from Eloundou et al. (2023). They create a new measure, observed exposure, to quantify which tasks that LLMs could theoretically speed up are actually seeing automated usage in professional settings. This measure captures several aspects of AI usage predictive of job impacts, such as whether tasks are theoretically possible with AI, see significant usage, and are performed in work-related contexts.

How exposure tracks with projected job growth and worker characteristics

The US Bureau of Labor Statistics (BLS) publishes regular employment projections. By comparing their predictions to the researchers' job-level coverage measure, they find that growth projections are somewhat weaker for jobs with more observed exposure. Workers in the top quartile of exposure tend to be different from those with zero exposure - they're more likely to be female, white, and Asian, earn more, and have higher levels of education. For example, people with graduate degrees are more common among the most exposed group.

Prioritizing outcomes

With exposure measures in hand, the question is what to look for. Researchers focus on unemployment as their priority outcome because it directly captures the potential for economic harm. They use the Current Population Survey to track unemployment trends, matching their occupation-level measures to respondents. A key question is which workers should be considered treated - should changes in employment be expected from just 10% task coverage?

Initial results

The researchers study trends in unemployment, comparing workers in the top quartile of time-weighted task coverage to those in the bottom. They find no impact on unemployment rates for workers in the most exposed occupations, although there's tentative evidence that hiring into those professions has slowed slightly for workers aged 22-25. The job finding rate for young workers in exposed occupations has decreased by about half a percentage point, with an averaged estimate of a 14% drop in the post-ChatGPT era.

Discussion

The report introduces a new measure for understanding AI's labor market effects and studies impacts on unemployment and hiring. The researchers find that computer programmers, customer service representatives, and financial analysts are among the most exposed occupations. While there's no impact on unemployment rates for workers in the most exposed occupations, there's tentative evidence that hiring into those professions has slowed slightly for young workers. The work is a first step toward cataloging AI's labor market impact, and the analytical steps taken can be updated as new data emerges.

Look, the reality is that AI's labor market impacts are still unfolding. The researchers' new measure provides a useful framework for understanding these effects. While there's no clear evidence of widespread job displacement yet, there are signs that hiring is slowing in exposed occupations, particularly for young workers. As AI continues to evolve, it's essential to keep monitoring its impact on the labor market and update our understanding accordingly.