Anthropic has published a working paper that finds no systematic increase in unemployment among workers in the occupations most exposed to generative AI since ChatGPT’s release, while also finding tentative evidence that hiring into those roles has slowed for workers ages 22 to 25.
A new exposure measure built from actual usage
The paper, written by Anthropic economists Maxim Massenkoff and Peter McCrory, introduces a new “observed exposure” measure built from real Claude usage data rather than theoretical model capability alone. Anthropic updated the paper on March 8 to correct Figure 7, which had reversed labels for job-start rates between the top-quartile and zero-exposure groups.
Anthropic presents the new measure as the paper’s main methodological contribution, saying it combines O*NET task data, usage data from the Anthropic Economic Index and task-level exposure estimates from Eloundou et al.’s 2023 “GPTs are GPTs” framework, which the paper describes as scoring whether an LLM could at least double task speed, with or without additional tools.
In Anthropic’s formulation, tasks count more heavily when they appear in work-related settings and in more automated, rather than merely assistive, patterns of use.
The gap between what AI can do and what it is doing
Anthropic highlights a gap between theoretical capability and observed deployment. Anthropic says large language models are theoretically capable of covering 94% of tasks in computer and math occupations and 90% in office and administrative roles under the Eloundou framework, yet Claude currently covers only 33% of tasks in the computer and math category.
In the occupation rankings Anthropic published, computer programmers lead observed exposure at 75% coverage, followed by customer service representatives, while data entry keyers are at 67%.
Using Current Population Survey data from August to October 2022, the paper says workers in the top quartile of exposure were 16 percentage points more likely to be female, 11 percentage points more likely to be white and almost twice as likely to be Asian as workers in the unexposed group.
The exposed group also earned 47% more on average, and 17.4% held graduate degrees versus 4.5% in the unexposed group.
No unemployment signal, but a narrower hiring result
On unemployment, Anthropic says trends for the most exposed and least exposed workers have been broadly similar since late 2022.
The paper argues that its framework should be able to detect a differential unemployment increase of about 1 percentage point, and says a doubling of unemployment in the top quartile of exposure, akin to a white-collar version of the Great Recession, should also be visible in its data if it had occurred. It did not find that signal.
The younger-worker result is narrower and more qualified. Anthropic says monthly job-start rates for workers ages 22 to 25 entering highly exposed occupations began to diverge in 2024, with entry into the most exposed jobs falling by about half a percentage point while job-finding rates in less exposed occupations remained around 2% a month.
The paper says that averages to a 14% drop in the post-ChatGPT era compared with 2022, but adds that the result is only barely statistically significant and may reflect other paths, including workers staying in existing jobs, taking different roles or returning to school.
That caveat is relevant because the Bureau of Labor Statistics defines unemployment counts only active job seekers as unemployed.
What the BLS projections say about the same occupations
Anthropic also ties its measure to official labor-market projections. The paper says occupations with higher observed exposure are projected by BLS to grow less through 2034, with each 10-point rise in coverage associated with a 0.6-point drop in projected growth.
BLS occupation pages also show mixed projections across some roles Anthropic flags: employment of computer programmers is projected to fall 6% from 2024 to 2034 and customer service representatives 5%, while the BLS table on the computer support specialists page shows computer user support specialists projected to decline 4%.
By contrast, medical records specialists are projected to grow 7%, even as BLS says wider adoption of AI-powered tools that make medical coding more efficient may affect demand.