Labor market impacts of AI: A new measure and early evidence
Citation. Massenkoff, M., & McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic.
Summary. Massenkoff and McCrory introduce “observed exposure,” a measure of AI displacement risk that combines O*NET tasks, theoretical LLM capability estimates, and Anthropic Economic Index usage data, weighting automated and work-related use more heavily than augmentative or non-work use. The report finds that observed exposure remains far below theoretical capability, correlates modestly with lower BLS occupational growth projections, and has not yet produced a systematic unemployment increase in highly exposed occupations. It does, however, report suggestive evidence that young workers are entering highly exposed occupations at lower rates after ChatGPT’s release.
Key claims
- Observed exposure is designed to measure which theoretically automatable tasks are actually seeing work-related AI use, then aggregate those task measures to occupations with weights for time spent on tasks and automation-like use.
- Observed Claude usage is concentrated in tasks that prior exposure measures classify as theoretically feasible: 97% of observed tasks fall into Eloundou et al.’s β=0.5 or β=1.0 categories, but realized coverage is still much lower than theoretical capability across occupational groups.
- Computer programmers have the highest observed exposure at 75% coverage, followed by customer service representatives and data entry keyers, while roughly 30% of workers are in occupations with zero observed coverage under the report’s threshold.
- A 10 percentage point increase in observed coverage is associated with a 0.6 percentage point lower BLS employment-growth projection for 2024–2034; the authors report no comparable correlation using the theoretical Eloundou et al. measure alone.
- Workers in the most exposed professions are more likely than unexposed workers to be female, white or Asian, highly educated, older, and higher paid; average earnings are 47% higher in the top-exposure group.
- Difference-in-differences estimates using Current Population Survey data show no clear post-ChatGPT unemployment increase for the most exposed occupations relative to unexposed occupations.
- For workers aged 22–25, the report estimates a roughly 14% drop in job-finding into highly exposed occupations compared with 2022, but treats this as barely statistically significant and open to alternative interpretations.
Evidence & limitations
- The report combines O*NET task data, task-level theoretical exposure estimates from Eloundou et al. (2023), Anthropic Economic Index usage data from August and November 2025, BLS employment projections, and Current Population Survey labor-market outcomes.
- The strongest contribution is measurement design: observed exposure separates theoretical capability from realized work-related usage and distinguishes automation-like from augmentation-like use before aggregating to occupations.
- Employment findings are early and intentionally cautious: the report finds no systematic unemployment effect through the available CPS window, while the young-worker hiring result is suggestive rather than definitive.
- The usage evidence comes from Claude/Anthropic data, so observed exposure may miss AI use in other tools, embedded enterprise systems, non-chat automation, or tasks avoided because of policy, risk, evaluation pressure, or lack of access.
- The inbox PDF hash differs from the public Anthropic CDN PDF hash, but extracted text, page count, and normalized text hash match exactly; the Drive hash is recorded in frontmatter and the public landing page is used as provenance.