Key findings
- Across the 50 states, a job-mix AI exposure score and measured AI use are positively related (correlation 0.46), but the job mix accounts for only 21% of the variation in use. Including DC lifts the correlation to 0.56, because DC is highest on both.
- States look alike on exposure and unlike on use. Exposure runs across just 0.51 points (Nevada to New York). Use runs across 15.5 points (West Virginia 20.8% to Maryland 36.3%). The trend line predicts only about 7.8 points of that from job mix.
- Nevada has the lowest job-mix exposure of any state (4.72) but 34.2% use, 9.0 points above its trend. Utah (+5.8) and Texas (+5.6) are next.
- West Virginia (-8.1), Maine (-8.0) and Vermont (-7.4) use AI furthest below what their job mix predicts.
- This does not show that occupations cause use. Age and urban or rural makeup are likely factors, and we could not test either (see below).
Two measures, built differently
AI use is Microsoft's estimate of the share of working-age residents who used a generative AI product, for Q1 2026. For states it is modelled from aggregated, anonymised Microsoft telemetry, not counted directly. We use the state table from Microsoft's US dataset.
Job-mix exposure is our occupation scores weighted by how many people each state employs in each occupation, from BLS Occupational Employment and Wage Statistics (OEWS), May 2025. It describes how much of a state's work looks like what AI tools can do. The scores are model-estimated, not measured, and this is the same occupation data behind our US analysis. We use OEWS only for each state's job-mix weights, never for totals, because the US country file is built on a different BLS product.
How exposure and use line up across states
The slope is real but states hardly spread out along it. Leaving out DC, every state sits between 4.7 and 5.2 on exposure, so almost all of the vertical movement in the chart comes from something other than occupations.
States that use AI more than their job mix predicts
The gap is a state's measured use minus the use the trend line predicts for its exposure score. These are the six largest gaps above the line among the 50 states.
| State | AI use, Q1 2026 | Use rank (of 51) | Job-mix exposure | Exposure rank (of 51) | Use the trend line predicts | Gap (points) |
|---|---|---|---|---|---|---|
| Nevada | 34.2% | 7 | 4.72 | 51 | 25.2% | +9.0 |
| Utah | 35.7% | 3 | 5.03 | 18 | 29.9% | +5.8 |
| Texas | 35.3% | 4 | 5.02 | 19 | 29.7% | +5.6 |
| Georgia | 33.7% | 10 | 5.00 | 25 | 29.3% | +4.4 |
| Florida | 33.6% | 11 | 5.01 | 21 | 29.5% | +4.1 |
| Illinois | 33.3% | 13 | 5.00 | 26 | 29.3% | +4.0 |
Source: site/data/us_states.json, built from Microsoft's state table and BLS OEWS (May 2025). Recomputed 2026-10-22.
Nevada's low score reflects its job mix. 13.2% of Nevada's OEWS jobs are in food preparation and serving, against 8.8% across all states, and 4.1% are in personal care and service, against 2.1%. Our model scores those groups low (3.5 out of 10). Only 1.9% of its jobs are computer and mathematical roles, against 3.4%. Yet Nevada's measured use is 34.2%. Our data does not say why. It does show that a job mix that looks low-exposure is no guarantee of low use.
States that use AI less than their job mix predicts
| State | AI use, Q1 2026 | Use rank (of 51) | Job-mix exposure | Exposure rank (of 51) | Use the trend line predicts | Gap (points) |
|---|---|---|---|---|---|---|
| West Virginia | 20.8% | 51 | 4.97 | 29 | 28.9% | -8.1 |
| Maine | 21.4% | 50 | 5.01 | 24 | 29.4% | -8.0 |
| Vermont | 23.3% | 47 | 5.09 | 11 | 30.7% | -7.4 |
| New Mexico | 23.9% | 45 | 5.05 | 16 | 30.1% | -6.2 |
| Montana | 22.7% | 49 | 4.95 | 32 | 28.6% | -5.9 |
| Alaska | 23.6% | 46 | 4.98 | 28 | 29.1% | -5.5 |
Source: same as above. A negative gap means measured use is below the trend line.
Vermont is the clearest case: its job mix ranks 11th of 51 on exposure, in the top quarter, while its use ranks 47th. West Virginia has the lowest use of any state (20.8%) and a below-average share of computer and mathematical jobs (2.1% against 3.4%). DC, Maryland and Utah, by contrast, rank first, second and third on use and all have above-average computer-job shares (6.1%, 5.1% and 4.2%), which is the pattern the positive correlation reflects.
What this does not show
- Age and place were not tested. Microsoft's own US report finds higher use in counties with many 18 to 24 year olds (28.6% average use against 20.3%) and in metropolitan areas. We could not add a state-level age control because the Census Bureau's data service needs an API key we do not have. Either factor could explain part or all of the 21% link.
- Three layers of estimates. Our exposure scores are model judgements, Microsoft's state use figures are modelled, and OEWS job counts are survey estimates.
- Small exposure spread. Excluding DC, state exposure scores differ by only 0.51 points, so rankings on that axis are sensitive to small changes in how occupations are scored.
- Incomplete occupation match. 49% to 59% of each state's jobs match one of our occupation codes exactly. The rest take the average score of our occupations in the same broad group, so every state has at least 93% of its jobs scored. The correlation is the same (0.56 across all 51) when only exact matches are used.
- Not a forecast. Neither measure predicts job losses.
What this means
If you work in a state with a high-exposure job mix, that tells you little about how widely AI is being used around you. The state-level gap in use is far larger than the gap in exposure, and the data here cannot say what explains it. If you work in a state with low measured use, the occupational exposure of your job is unchanged: the tools may simply have reached fewer people so far.
For analysts, the useful reading is the outliers. A state far from the trend line is a place where something other than occupation is shaping use, which is where a closer look, with age and local economy data, would start.
Check any country's full occupation breakdown
See AI exposure scores broken down by every occupation group, for over 200 countries, using official ILO and national statistics data.
Open the United States in the explorer →AI exposure, robotics, offshoring and work-from-home scores are model-estimated: an AI language model (Claude Opus 4.7) scored each ISCO-08 occupation group on 28 May 2026, guided by published research. They are informed judgements, not measurements, and not forecasts of job losses. Read the full methodology and how the scores were checked.
Frequently asked questions
Do US states with more AI-exposed job mixes use AI more?
Which US states use AI more than their job mix predicts?
Which US states use AI less than their job mix predicts?
Where does the state AI use data come from?
Get new country analyses, data updates, and AI labour market insights. No spam - one email when something worth reading drops.
Sources
- Microsoft AI Economy Institute (2026) - AI Diffusion in the United States, Q1 2026 data, MIT licence
- Misra et al. (2026) - Measuring AI Diffusion Across U.S. Geographies: County, State, and Metro Estimates
- US Bureau of Labor Statistics - Occupational Employment and Wage Statistics, May 2025, state estimates
- WorldJobsData (2026) - Occupation-level AI exposure scores, methodology
Spotted an error or know a better source? Report it and we will check it against the source data.