AI Job Displacement by County: Where the Risk Is Concentrating

Why a county view matters

National headlines about AI and jobs almost always quote a single number: the share of US tasks "exposed" to generative AI. The figure most often cited comes from Eloundou and colleagues at OpenAI, who estimated in 2023 that roughly 80 percent of the US workforce could have at least 10 percent of their tasks affected by large language models, with about 19 percent of workers seeing at least half of their tasks exposed (Eloundou et al., 2023). That headline is real, but it hides the more useful question for workers, mayors, and economic developers: where, geographically, does the exposure land first?

County-level analysis answers that question. The labor market is not a national pool. It is roughly 3,140 county-level pools with different occupational mixes, different industry concentrations, and different starting wages. An AI capability that automates legal document review hits Manhattan, Washington DC, and Cook County, Illinois, very differently from how it hits a rural Appalachian county whose largest employers are a hospital and a school district.

How exposure is actually measured

The credible exposure estimates all use a similar pipeline. Start with the O*NET task database maintained by the US Department of Labor, which decomposes each occupation into the discrete tasks workers actually perform. Score each task for whether a current AI system can perform it at human-equivalent quality. Aggregate task scores up to occupation scores. Then take the Bureau of Labor Statistics Occupational Employment and Wage Statistics (OEWS) tables, which report how many workers in each occupation are employed in each metro area, and weight the exposure score by local employment.

The most-cited county-level work in this style comes from Felten, Raj, and Seamans (2021), updated for generative AI in 2023 (NBER working paper). Brookings Metro has applied a similar approach to map AI exposure for every US metropolitan area in its The Geography of Generative AI series, led by Mark Muro and Yang You. The Stanford 2024 AI Index chapter on the labor market is a useful synthesis of where these estimates converge.

Where exposure clusters

Three patterns are now well-documented.

First, high-wage knowledge-work counties show the highest task exposure. Counties with dense concentrations of lawyers, financial analysts, software developers, marketing professionals, and writers, places like New York County NY, San Francisco County CA, Arlington County VA, and Suffolk County MA, score in the top decile of generative-AI exposure. This inverts the older automation literature, which found risk concentrated in routine manual work.

Second, back-office service counties show high exposure with lower wages. Counties whose economies are anchored by call centers, claims processing, payroll, and bookkeeping operations, parts of Maricopa County AZ, Tarrant County TX, and many suburban Atlanta and Charlotte counties, have exposure scores near the top of the distribution and median wages near the bottom of the exposed group. The harm if displacement happens is larger here per worker because reskilling pathways are thinner.

Third, rural counties dominated by health care, education, construction, agriculture, and personal services have low direct exposure. The Acemoglu and Restrepo research program at MIT (Acemoglu, 2024) suggests that low direct exposure does not translate into immunity. Indirect effects, through demand from exposed sectors, supply chains, and migration, propagate.

Exposure is not displacement

It is worth being clear about what the exposure score is and is not. It is a measure of technical feasibility: could a current AI system, in principle, do this task. It is not a forecast of who will lose a job, or when, or with what replacement income. The translation from exposure to actual employment change depends on adoption speed, capital costs, regulatory friction, labor-market institutions, and management choices that no model captures cleanly.

The OECD reviewed the early evidence in its 2023 Employment Outlook and concluded that "few jobs have been lost to AI so far," while warning that the speed of generative-AI deployment is unlike previous automation waves. Goldman Sachs Research has estimated that generative AI could affect the equivalent of 300 million full-time jobs globally, with the caveat that historical productivity gains have always been accompanied by new job creation (Briggs and Kodnani, 2023). The honest reading is that exposure scores tell you who is at the front of the line, not who is being shown the door this year.

What this means for state and local policy

If a county economic development office wants to use this data well, three operational moves follow.

• Identify the top three exposed occupations by local employment, not by national rank. A county with 4,000 medical secretaries faces a different conversation than a county with 4,000 paralegals.
• Map wage distributions inside exposed occupations. A 20 percent productivity gain that shows up as wage growth in high-skill exposed roles can show up as headcount reduction in low-wage exposed roles.
• Inventory the local training infrastructure, community colleges, apprenticeship intermediaries, and Workforce Innovation and Opportunity Act (WIOA) providers, against the occupations workers can plausibly transition into. Reskilling capacity is the binding constraint, not exposure.

These data choices connect to broader regional questions covered in our companion pieces: state-level adoption capacity in AI readiness by state, affordability under wage compression in the American Dream Affordability Index, and structural mobility differences across counties in our Geography of Opportunity Rankings.

Methodological caveats worth printing on the report

Three caveats should appear on any map of AI exposure that a policymaker reads.

The exposure score is a snapshot of current model capability. A 2026 score will not match a 2028 score. Scores are most stable for tasks that involve standardized inputs and verifiable outputs, and least stable for tasks that require sustained judgement, negotiation, or physical presence.

The OEWS occupation counts are estimates, not censuses. Small metros have larger confidence intervals. County-level disaggregation, where it is published, inherits that uncertainty.

And exposure is not the same as harm. A county that scores high on exposure may capture most of the productivity gain in wages, in lower prices for consumers, or in new business formation. The distributional question, who wins and who loses, is determined by the rules around adoption, not by the technology score itself. The foundational reasoning behind these claims is laid out in our foundations explainer on what AI actually is and the practical limits of model accuracy in Large Language Models, Explained Without the Hype.

Where to go next

Readers who want the underlying microdata can pull the BLS OEWS tables for every metro area and merge them against the published exposure scores. The Brookings Metro interactive map is the most accessible visualization. For the labor-economics literature behind the numbers, the NBER labor studies working paper series carries new exposure papers most months.

The takeaway is not that AI will or will not displace workers in your county. It is that the question can now be asked at a county level, and that the answer is structured enough to plan against. The places that will weather the next decade well are the ones that take the exposure map seriously without mistaking it for prophecy.

Sources

• Eloundou, Manning, Mishkin and Rock, *GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models*, 2023.
• Felten, Raj and Seamans, *Occupational, industry, and geographic exposure to artificial intelligence*, Strategic Management Journal, 2021; updated in NBER w31178.
• Brookings Metro, *The Geography of Generative AI*, 2023.
• OECD, *Employment Outlook 2023: Artificial Intelligence and the Labour Market*.
• Acemoglu, *The Simple Macroeconomics of AI*, NBER 2024.
• Stanford HAI, *2024 AI Index Report*.
• Bureau of Labor Statistics, Occupational Employment and Wage Statistics.
• US Department of Labor, O*NET Online Task Database.