Anthropic maps AI job risk as most exposed workers are female and educated

Anthropic economists MMMaxim Massenkoff and PMPeter McCrory published Labor Market Impacts of AI: A New Measure and Early Evidence on March 5, 2026. The paper introduces a metric called observed exposure that distinguishes between what AI is theoretically capable of doing and what AI is actually doing in professional settings. The methodology combines data on which tasks large language models are capable of performing with anonymized, real-world usage data from the Claude model. The researchers weighted automated uses — where AI replaces human work entirely — more heavily than augmentative uses where AI assists. The most exposed occupations by Claude usage data are computer programmers at 75% task coverage, customer service representatives at 70.1%, data entry keyers at 67.1%, and medical records specialists at a similarly elevated level. At the bottom — zero percent — are cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants.

The demographic profile of AI-exposed workers is the paper''s most politically significant finding. Workers in the top quartile of AI-exposed occupations are 16 percentage points more likely to be female than workers in the least-exposed group. They earn 47% more than the least-exposed group. They are nearly four times as likely to hold a graduate degree. This profile — educated, higher-earning, female-skewing — is the opposite of the workers who faced displacement from manufacturing offshoring or previous waves of automation, which primarily hit lower-income, lower-education, male-skewing workers. This demographic reversal was made politically explicit one week later when Palantir CEO Alex Karp told CNBC that AI disrupts humanities-trained, largely Democratic voters and makes their economic power less.

Despite the high observed exposure figures for white-collar occupations, the study finds no statistically significant increase in unemployment among workers in those occupations as of early 2026. The unemployment rate for highly exposed workers has not risen relative to less-exposed workers since the release of ChatGPT in late 2022. The researchers caution that this non-finding does not mean displacement will not happen — it means the effects may be slow and structurally hidden, similar to how the China shock of the early 2000s took a decade to fully appear in economic data despite being severe.

The paper explicitly warns about a scenario it names the Great Recession for white-collar workers. During the 2007–2009 financial crisis, U.S. unemployment doubled from approximately 5% to 10% in less than two years. If something comparable happened among the top quartile of AI-exposed occupations — where the current unemployment rate is around 3% — a doubling to 6% would represent millions of displaced workers concentrated in the educated, professional class. The researchers note that such a scenario has not happened yet but absolutely could.

The one concrete early signal the paper identifies is in entry-level hiring. Among workers aged 22 to 25, hiring into AI-exposed occupations has slowed measurably. A separate referenced study found a 16% fall in employment in AI-exposed jobs among this age group. Young workers who are not being hired may be staying in lower-exposure roles, taking unrelated jobs, or returning to school. This quiet displacement — where jobs technically exist but entry points are closing — is exactly the dynamic that existing federal data systems are not designed to capture.

The gap between AI''s theoretical capability and its actual observed use is central to interpreting the research. In computer and mathematical occupations, AI could theoretically handle 94% of tasks — but actual observed usage covers only about 33%. The researchers attribute this gap to legal constraints that prevent AI from acting autonomously in regulated contexts, the need for additional software integration, continued human verification requirements, and slower organizational adoption cycles than individual usage. This gap is closing continuously as infrastructure, legal frameworks, and organizational readiness advance.

The Census Bureau adjusted how it surveys businesses about AI usage in late 2025, and the methodology change resulted in a sharp increase in the share of firms reporting AI use — meaning prior estimates of AI adoption understated actual deployment significantly. On March 6, 2026, nine bipartisan senators — including MWMark Warner, Josh Hawley, and Todd Young — sent a letter to Labor Secretary Lori Chavez-DeRemer demanding that BLS and Census expand their surveys to track AI''s workforce impact. The senators cited Anthropic''s research directly as evidence that private companies are publishing more credible AI labor market data than the federal government.