Anthropic’s AI Labor Impact Report Reveals Surprising Employment Gaps

New Measurement Standard

Anthropic introduced the “exposure” metric, which combines theoretical AI capability scores with real‑world usage data from Claude to measure the proportion of a job’s tasks actually performed by AI.

Data sources: O*NET database covering ~800 U.S. occupations, Claude usage logs, and task‑level theoretical capability assessments. 97 % of tasks observed in Claude belong to categories that are theoretically feasible.

Example: computer‑and‑math occupations have a theoretical coverage of 94 % but an observed coverage of only 33 %, attributed to legal constraints, specialized software requirements, and manual verification steps.

High‑Risk Occupation Rankings

Computer programmers rank highest with a 75 % exposure rate, followed by customer‑service representatives and data‑entry clerks at 67 %.

Since ChatGPT’s launch, entry‑level hiring for these high‑risk occupations fell 14 %; recent graduates are affected about four times more than other workers.

Who Is Most Affected?

Workers in high‑exposure occupations are more likely to be women, white, older, and higher‑earning. Individuals with graduate degrees are nearly four times as prevalent in the high‑exposure group compared to the low‑exposure group.

Safe Havens

Manual jobs such as bartenders, dishwashers, and lifeguards together account for about 30 % of the labor market and show zero exposure because their tasks are not currently replaceable by AI.

Overall, 30 % of workers show zero exposure because their tasks appear too infrequently in the data to meet the minimum threshold.

Prediction Validation

Comparing exposure with the U.S. Bureau of Labor Statistics 2024‑2034 employment‑growth forecasts shows that each 10‑point increase in exposure corresponds to a 0.6‑point decrease in projected growth.

This correlation disappears when using only the theoretical capability metric, highlighting the importance of real‑world usage data.

Employment Data Analysis

Unemployment trends since ChatGPT’s release show no systematic rise for high‑exposure jobs, but the proportion of 22‑25‑year‑old workers entering these roles has declined.

The impact appears as slower hiring rather than increased layoffs; many young workers remain in current positions, shift to other occupations, or return to education.

Research Limitations

The report notes lag in the data and likely underestimation of impact. Future work will incorporate more usage data, update capability assessments (currently based on early‑2023 LLMs), and examine young workers and recent graduates in greater depth.

Full report available at https://www.anthropic.com/research/labor-market-impacts