Jobs Most Exposed to AI (2026): Exposure vs Measured Job Losses

Computer programmers, data entry keyers, customer service representatives and financial analysts are the occupations most exposed to AI by the newest measure, Anthropic's March 2026 analysis of what AI systems are actually used for against the tasks each job contains. Programmers' tasks are 75% covered, data entry keyers' 67%. By the older capability-based measures from 2023, 19% of US workers are in the most exposed jobs, and they are disproportionately college-educated, female and well-paid. Exposure is not job loss. As of the latest data, measured unemployment effects are small and concentrated among workers aged 22 to 25 in exposed occupations.

The lists that rank for this question lead with a 2023 estimate that 300 million jobs could be replaced, which the study did not say, or with a World Economic Forum forecast repeated as a measurement. They do not distinguish a model being able to do a task from anyone using it to, and they cite employment projections that the Bureau of Labor Statistics replaced on August 27, 2026. This page separates the three questions: what could AI do, what is it being used for, and what has happened to jobs so far. Every figure has its source, sample and date.

Three ways to measure exposure

Every ranking of AI-exposed jobs comes from one of three methods, and they give different lists.

• Capability exposure asks whether a model could do each task in a job, using the Department of Labor's O*NET task database. Eloundou, Manning, Mishkin and Rock's "GPTs are GPTs" (March 2023) found about 80% of US workers could have at least 10% of their tasks affected by large language models and 19% could have at least 50% affected. Pew Research Center's July 26, 2023 analysis applied a similar method to 2022 Current Population Survey data. Felten, Raj and Seamans' AI Occupational Exposure index (2021, updated 2023) links AI capabilities to the abilities each job needs.
• Observed exposure asks what people actually use AI for. Anthropic's Economic Index (September 15, 2025) classified a million Claude conversations and a million API transcripts by O*NET task; its March 5, 2026 paper "Labor market impacts of AI" combined that usage data with capability scores into an "observed exposure" measure across about 800 occupations, giving full weight to tasks AI automates and half weight to tasks it assists.
• Skills exposure asks how many of a job's skills AI can perform. Indeed Hiring Lab's AI at Work report (September 23, 2025) scored 2,884 skills across US job postings.

Capability measures from 2023 rank jobs by what GPT-4 might do. Observed measures from 2026 rank them by what people did with Claude in August and November 2025. The gap between the two is large: Anthropic found AI could theoretically cover 94% of computer and mathematical tasks and was actually used for about 33% of them, and about 30% of US workers had zero observed coverage.

The most exposed occupations by each measure

The lists agree on the shape: jobs built from writing, coding, calculating and processing text are exposed; jobs built from physical presence and physical skill are not. They disagree on the order, and the 2023 lists include jobs such as drafters and budget analysts that the 2026 usage data do not show being automated at scale.

Who holds the exposed jobs

Pew's 2023 analysis is the clearest demographic breakdown, because it uses the Current Population Survey rather than platform data. Of US workers in the most exposed jobs: 27% of those with a bachelor's degree or more against 12% with a high school education only; 21% of women against 17% of men; 24% of Asian workers, 20% of white, 15% of Black and 13% of Hispanic workers. Workers in the most exposed jobs earned $33 an hour on average against $20 in the least exposed. Brookings' October 2024 analysis of 1,016 occupations found 36% of women against 25% of men in jobs where generative AI could save half their task time, with office and administrative support, employing nearly 19 million people, the largest exposed group. This is the inversion from earlier automation, which fell on lower-paid, male-dominated manual work.

What the 2025 to 2035 BLS projections say

The Bureau of Labor Statistics released its 2025 to 2035 employment projections on August 27, 2026, replacing the 2024 to 2034 edition that most pages still cite. The release names generative AI as a factor that may limit jobs in arts, design, entertainment and media, and says automation including AI tools will reduce demand for office and administrative support, which is projected to lose 752,100 jobs, a 4.0% decline, the largest of any group. Sales occupations are projected at -1.4% and production at -0.4%.

The table makes the central point. The occupation with the highest observed exposure, software development, is projected to grow 10%. The occupations projected to shrink fastest, data entry and customer service, were shrinking before generative AI. Anthropic's paper found that a ten-point increase in its coverage measure was associated with 0.6 points lower projected growth in the BLS series, a real but modest relationship.

What has been measured so far

• Unemployment, all exposed workers. Anthropic's March 2026 analysis of Current Population Survey data found no systematic increase in unemployment for workers in the top exposure quartile relative to those with zero exposure since late 2022, with a design able to detect a change of about one percentage point. The St. Louis Fed (Ozkan and Sullivan, August 26, 2025) found a 0.57 correlation between occupational exposure and the rise in unemployment from 2022 to 2025, with computer and mathematical occupations up 1.2 points; the authors called the result tentative and not causal.
• Hiring of young workers. Stanford Digital Economy Lab's "Canaries in the Coal Mine", using ADP payroll data, found in August 2025 that employment of workers aged 22 to 25 in the most exposed occupations was 13% lower than it would have been relative to less exposed peers; a revision with data through June 2026, published August 12, 2026, put the gap at 19%. The gap comes from lower hiring, not more separations, and experienced workers show no comparable effect. Anthropic found the monthly rate at which 22 to 25 year olds found jobs in exposed occupations was 14% lower than in 2022, described by the authors as "just barely statistically significant", with no decline for over-25s.
• Occupational mix. Yale Budget Lab's tracker, updated with March 2026 data, found the change in the occupational mix of the US workforce since ChatGPT's release within historical ranges and slightly predating it, with one uptick in the gap between recent college graduates aged 20 to 24 and those aged 25 to 34.

The consistent reading across the three groups is: no broad job loss yet, a real and growing entry-level hiring effect in exposed occupations, and no effect on experienced workers in the same jobs. Our AI in the workplace statistics page covers the productivity experiments that sit behind these numbers.

Layoffs attributed to AI

Challenger, Gray and Christmas has tracked AI as a stated reason for announced job cuts since 2023. Its July 2026 report counted 10,970 cuts attributed to AI that month, 33% of all announced cuts and the fifth consecutive month AI was the most cited reason; 112,713 for January to July 2026, about 24% of the year's total; and 184,538 since tracking began. July's total announced cuts, 33,429, were the lowest monthly figure in two years. For scale, the US labor force is about 170 million and total announced cuts in 2025 were about 1.2 million, of which AI was cited for 54,836, or 5%. Challenger's count records what employers said, and its own report notes that naming AI can please investors. It is the best series on the subject and should be read as attribution, not cause.

The least exposed jobs

Pew's least exposed occupations were nannies, barbers and child care workers; 23% of all workers were in its least exposed group. Indeed's least transformed occupation was nursing, with 68% of its skills in the minimal tier. Anthropic found about 30% of workers with zero observed coverage. The pattern is physical presence: construction trades, installation and repair, health care support, food preparation, personal care and transport. The BLS projects health care support and health care practitioner occupations as the fastest growing groups to 2035, and those are also the least exposed.

Figures to stop using

• "300 million jobs will be lost to AI." Goldman Sachs' March 2023 report said the equivalent of 300 million full-time jobs globally were exposed to automation, that about two thirds of US jobs were exposed to some degree, that 25% to 50% of the workload in exposed jobs could be replaced, and that 63% of US workers would be complemented rather than substituted, with 7% substituted. Exposed is not lost.
• "83 million jobs by 2027" or "92 million displaced." World Economic Forum figures from employer surveys about expected change, repeated as measurements. The same reports project more jobs created than displaced.
• "80% of workers will be affected." Eloundou's 80% is workers with at least 10% of tasks exposed. The 50%-of-tasks figure is 19%.
• The 2023 occupation lists as current. They measured GPT-4's possible capabilities. Observed usage in 2025 and 2026 covers about a third of what was possible.
• Claude usage shares as employment shares. 36% of Claude conversations being coding does not mean 36% of exposure is in software. It describes the users of one product.
• The Stanford 13% without the revision. The August 2026 update is 19%. Cite both with dates.
• Anthropic's 14% as percentage points. It is a 14% relative decline in a monthly job-finding rate, and barely significant.
• The 2024 to 2034 BLS projections. Superseded on August 27, 2026.

Key takeaways

• By observed use, computer programmers (75% of tasks), data entry keyers (67%), customer service representatives and financial analysts are the most exposed occupations (Anthropic, March 5, 2026).
• By 2023 capability measures, 19% of US workers are in the most exposed jobs, 80% have at least 10% of tasks exposed, and exposure is highest among college-educated (27% vs 12%), female (21% vs 17%) and higher-paid ($33 vs $20 an hour) workers.
• AI could theoretically cover 94% of computer and mathematical tasks; observed use covers about 33%. About 30% of US workers have zero observed coverage.
• BLS projections for 2025 to 2035, released August 27, 2026, show office and administrative support losing 752,100 jobs (-4.0%), data entry keyers -25.5% and customer service -5.3%, while software developers grow 10%.
• No systematic rise in unemployment for exposed workers has been measured. Workers aged 22 to 25 in exposed jobs show 19% lower employment than peers (Stanford/ADP, August 2026, revised from 13%) and a 14% lower job-finding rate (Anthropic), driven by hiring not separations.
• Employers attributed 112,713 announced cuts to AI in January to July 2026, 24% of all cuts, and 184,538 since 2023 (Challenger). The 2025 full-year share was 5%.
• Goldman Sachs' 300 million figure is jobs exposed, not lost; the same report expected 63% of US workers to be complemented and 7% substituted.
• Least exposed: nannies, barbers, child care workers, nurses and the physical trades, which are also the BLS's fastest growing groups to 2035.