Anthropic proposes 'observed exposure' metric for AI labor impact and finds limited unemployment effects so far

Anthropic published a new economic research paper on March 5, 2026 that tries to measure labor disruption from AI using a metric it calls observed exposure. Instead of looking only at what LLMs could theoretically do, the paper combines theoretical task capability with real-world Claude usage and gives more weight to automated, work-related activity.

The paper uses O*NET task definitions, Anthropic Economic Index data, and prior task-level exposure estimates to ask a practical question: where is AI actually showing up in professional work? Anthropic says the answer is still far below the technology's theoretical ceiling. In its data, computer and math jobs have high potential exposure, but current Claude coverage remains a much smaller fraction of all tasks than pure capability estimates suggest.

• Anthropic says occupations with higher observed exposure are projected by the U.S. Bureau of Labor Statistics to grow less through 2034.
• The most exposed workers are more likely to be older, female, more educated, and higher paid.
• The paper finds no systematic increase in unemployment for highly exposed workers since late 2022.
• It does find tentative evidence that hiring for younger workers has slowed in more exposed occupations.

Anthropic highlights computer programmers, customer service representatives, and data entry roles among the most exposed jobs in its framework. At the other end, many physical or in-person occupations still show effectively zero coverage in the company's usage data. That gap matters because it suggests diffusion, workflow design, regulation, and human verification remain major constraints even when models can theoretically perform a task.

The practical value of the paper is not a single headline number. It offers a repeatable framework for checking whether AI adoption is translating into hiring or unemployment effects over time. For companies, policymakers, and workers, that is more useful than broad claims that AI is either already replacing labor everywhere or having no measurable impact at all.