Labor market impacts of AI: A new measure and early evidence

This report, published by Anthropic in March 2026 and written by researchers Maxim Massenkoff and Peter McCrory, presents a new methodological framework to analyze how artificial intelligence is transforming the labor market in the United States.

The starting point is a critique of previous approaches: studies that predicted major labor disruptions —such as those identifying 25% of jobs as vulnerable to offshoring— did not materialize in practice. The authors argue that measuring AI's impact is especially difficult because its effects are gradual and diffuse, more similar to those of the internet or trade with China than to a sudden crisis like the COVID-19 pandemic.

To address these limitations, the study introduces a new measure called Observed Exposure, combining three data sources: the O*NET occupational task database, theoretical LLM capability estimates from Eloundou et al. (2023), and real Claude usage data from the Anthropic Economic Index. Unlike purely theoretical measures, this metric weights automated uses —over assistive ones— and professional work contexts more heavily. The result is a more realistic picture of which jobs are actually being affected by AI today, not just which ones could theoretically be.

Among the key findings, the document shows a significant gap between what AI can theoretically do and what it actually does: observed coverage is only a fraction of theoretical capability. The most exposed occupations are computer programmers (74.5%), customer service representatives (70.1%), and data entry operators (67.1%). At the other end, 30% of workers have zero recorded exposure, including cooks, mechanics, and bartenders.

The most exposed workers tend to be women, over 40, with higher education and higher salaries. Bureau of Labor Statistics projections indicate that the most exposed occupations will grow less over the next decade.

Regarding current impact, the analysis finds no evidence of a systematic increase in unemployment among the most exposed workers since late 2022. However, it does detect an early signal: hiring of young workers (22–25) in exposed occupations has fallen approximately 14% since 2022, though the result is barely statistically significant.

The report is aimed at labor economics researchers, policymakers, technology sector professionals, and anyone interested in understanding, with empirical rigor, the real —and still limited— effects of AI on employment.

Key points

• New metric Observed Exposure: combines theoretical LLM capability with real Claude usage
• Large gap between what AI can do and what it actually does
• Most exposed occupations: programmers (74.5%), customer service (70.1%), data entry (67.1%)
• 30% of workers have no recorded exposure
• Most exposed workers: women, over 40, higher education, higher salaries
• BLS projects lower employment growth in exposed occupations through 2034
• No evidence of increased unemployment among the most exposed since 2022
• Early signal: hiring of young workers (22–25) in exposed jobs fell 14%